enforce model price calculation

add support to using custom model ids

migrations/versions/b1c2d3e4f5a6_add_upstream_model_id_to_models.py

@@ -0,0 +1,33 @@
+"""add forwarded_model_id to models
+
+Revision ID: b1c2d3e4f5a6
+Revises: a776ca70e5fe
+Create Date: 2026-04-05 00:00:00.000000
+"""
+
+import sqlalchemy as sa
+import sqlmodel
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "b1c2d3e4f5a6"
+down_revision = "a776ca70e5fe"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "models",
+        sa.Column(
+            "forwarded_model_id",
+            sqlmodel.sql.sqltypes.AutoString(),
+            nullable=True,
+        ),
+    )
+    # Backfill: set forwarded_model_id = id for all existing rows
+    op.execute("UPDATE models SET forwarded_model_id = id WHERE forwarded_model_id IS NULL")
+
+
+def downgrade() -> None:
+    op.drop_column("models", "forwarded_model_id")

routstr/algorithm.py

@@ -217,6 +217,10 @@ def create_model_mappings(
                 if prefixed_id not in aliases:
                     aliases.append(prefixed_id)
 
+            # Register forwarded_model_id as a routable alias
+            if model_to_use.forwarded_model_id and model_to_use.forwarded_model_id not in aliases:
+                aliases.append(model_to_use.forwarded_model_id)
+
             # Try to set each alias
             for alias in aliases:
                 _add_candidate(alias, model_to_use, upstream)
@@ -305,6 +309,10 @@ def create_model_mappings(
             if prefixed_id not in aliases:
                 aliases.append(prefixed_id)
 
+        # Register forwarded_model_id as a routable alias
+        if model_to_use.forwarded_model_id and model_to_use.forwarded_model_id not in aliases:
+            aliases.append(model_to_use.forwarded_model_id)
+
         for alias in aliases:
             _add_candidate(alias, model_to_use, upstream_for_override)
         seen_model_provider.add(dedupe_key)

routstr/auth.py

@@ -7,6 +7,7 @@ from datetime import datetime
 from typing import Optional
 
 from fastapi import HTTPException
+from sqlalchemy import case
 from sqlalchemy.exc import IntegrityError
 from sqlmodel import col, select, update
 
@@ -22,6 +23,7 @@ from .payment.cost_calculation import (
 from .wallet import credit_balance, deserialize_token_from_string
 
 logger = get_logger(__name__)
+payments_logger = get_logger("routstr.payments")
 
 # TODO: implement prepaid api key (not like it was before)
 # PREPAID_API_KEY = os.environ.get("PREPAID_API_KEY", None)
@@ -584,6 +586,18 @@ async def pay_for_request(
             "total_requests": billing_key.total_requests,
         },
     )
+    payments_logger.info(
+        "RESERVE",
+        extra={
+            "event": "reserve",
+            "key_hash": key.hashed_key[:8] + "...",
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+            "cost_reserved": cost_per_request,
+            "balance": billing_key.balance,
+            "reserved_balance": billing_key.reserved_balance,
+            "total_spent": billing_key.total_spent,
+        },
+    )
 
     return cost_per_request
 
@@ -635,6 +649,17 @@ async def revert_pay_for_request(
     await session.refresh(billing_key)
     if billing_key.hashed_key != key.hashed_key:
         await session.refresh(key)
+    payments_logger.info(
+        "REVERT",
+        extra={
+            "event": "revert",
+            "key_hash": key.hashed_key[:8] + "...",
+            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+            "cost_reverted": cost_per_request,
+            "balance": billing_key.balance,
+            "reserved_balance": billing_key.reserved_balance,
+        },
+    )
     return True
 
 
@@ -728,11 +753,32 @@ async def adjust_payment_for_tokens(
                 },
             )
             # Finalize by releasing reservation and charging max cost
+            if billing_key.reserved_balance < deducted_max_cost:
+                logger.error(
+                    "reserved_balance below deducted_max_cost before MaxCost finalization — clamping to 0",
+                    extra={
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "reserved_balance": billing_key.reserved_balance,
+                        "deducted_max_cost": deducted_max_cost,
+                        "total_cost_msats": cost.total_msats,
+                        "balance": billing_key.balance,
+                        "total_spent": billing_key.total_spent,
+                        "model": model,
+                    },
+                )
+
+            safe_reserved = case(
+                (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                 col(ApiKey.reserved_balance) - deducted_max_cost),
+                else_=0,
+            )
+
             finalize_stmt = (
                 update(ApiKey)
                 .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                 .values(
-                    reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                    reserved_balance=safe_reserved,
                     balance=col(ApiKey.balance) - cost.total_msats,
                     total_spent=col(ApiKey.total_spent) + cost.total_msats,
                 )
@@ -741,13 +787,17 @@ async def adjust_payment_for_tokens(
 
             # Also update total_spent and reserved_balance on the child key if it's different
             if billing_key.hashed_key != key.hashed_key:
+                child_safe_reserved = case(
+                    (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                     col(ApiKey.reserved_balance) - deducted_max_cost),
+                    else_=0,
+                )
                 child_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == key.hashed_key)
                     .values(
                         total_spent=col(ApiKey.total_spent) + cost.total_msats,
-                        reserved_balance=col(ApiKey.reserved_balance)
-                        - deducted_max_cost,
+                        reserved_balance=child_safe_reserved,
                     )
                 )
                 await session.exec(child_stmt)  # type: ignore[call-overload]
@@ -782,6 +832,23 @@ async def adjust_payment_for_tokens(
                         "model": model,
                     },
                 )
+                payments_logger.info(
+                    "FINALIZE",
+                    extra={
+                        "event": "finalize",
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "model": model,
+                        "cost_reserved": deducted_max_cost,
+                        "cost_charged": cost.total_msats,
+                        "input_tokens": cost.input_tokens,
+                        "output_tokens": cost.output_tokens,
+                        "balance": billing_key.balance,
+                        "reserved_balance": billing_key.reserved_balance,
+                        "total_spent": billing_key.total_spent,
+                        "finalize_type": "max_cost",
+                    },
+                )
             return cost.dict()
 
         case CostData() as cost:
@@ -815,12 +882,32 @@ async def adjust_payment_for_tokens(
                         "model": model,
                     },
                 )
+                if billing_key.reserved_balance < deducted_max_cost:
+                    logger.error(
+                        "reserved_balance below deducted_max_cost on exact-cost finalization — clamping to 0",
+                        extra={
+                            "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                            "reserved_balance": billing_key.reserved_balance,
+                            "deducted_max_cost": deducted_max_cost,
+                            "total_cost_msats": total_cost_msats,
+                            "balance": billing_key.balance,
+                            "total_spent": billing_key.total_spent,
+                            "model": model,
+                        },
+                    )
+
+                exact_safe_reserved = case(
+                    (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                     col(ApiKey.reserved_balance) - deducted_max_cost),
+                    else_=0,
+                )
+
                 finalize_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance)
-                        - deducted_max_cost,
+                        reserved_balance=exact_safe_reserved,
                         balance=col(ApiKey.balance) - total_cost_msats,
                         total_spent=col(ApiKey.total_spent) + total_cost_msats,
                     )
@@ -829,13 +916,17 @@ async def adjust_payment_for_tokens(
 
                 # Also update total_spent and reserved_balance on the child key if it's different
                 if billing_key.hashed_key != key.hashed_key:
+                    child_exact_safe_reserved = case(
+                        (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                         col(ApiKey.reserved_balance) - deducted_max_cost),
+                        else_=0,
+                    )
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
                         .values(
                             total_spent=col(ApiKey.total_spent) + total_cost_msats,
-                            reserved_balance=col(ApiKey.reserved_balance)
-                            - deducted_max_cost,
+                            reserved_balance=child_exact_safe_reserved,
                         )
                     )
                     await session.exec(child_stmt)  # type: ignore[call-overload]
@@ -844,44 +935,55 @@ async def adjust_payment_for_tokens(
                 await session.refresh(billing_key)
                 if billing_key.hashed_key != key.hashed_key:
                     await session.refresh(key)
-                return cost.dict()
-
-            # this should never happen why do we handle this???
-            if cost_difference > 0:
-                # Need to charge more than reserved, finalize by releasing reservation and charging total
-                logger.info(
-                    "Additional charge required for token usage",
+                payments_logger.info(
+                    "FINALIZE",
                     extra={
+                        "event": "finalize",
                         "key_hash": key.hashed_key[:8] + "...",
                         "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                        "additional_charge": cost_difference,
-                        "current_balance": billing_key.balance,
-                        "sufficient_balance": billing_key.balance >= cost_difference,
                         "model": model,
+                        "cost_reserved": deducted_max_cost,
+                        "cost_charged": total_cost_msats,
+                        "input_tokens": cost.input_tokens,
+                        "output_tokens": cost.output_tokens,
+                        "balance": billing_key.balance,
+                        "reserved_balance": billing_key.reserved_balance,
+                        "total_spent": billing_key.total_spent,
+                        "finalize_type": "exact",
                     },
                 )
+                return cost.dict()
+
+            # actual cost exceeded discounted reservation (due to tolerance_percentage)
+            if cost_difference > 0:
+                # Always release the reservation and charge min(actual_cost, balance).
+                # Using a CASE expression makes this a single atomic UPDATE — no
+                # multi-level fallback needed and balance can never go negative.
+                chargeable = case(
+                    (col(ApiKey.balance) >= total_cost_msats, total_cost_msats),
+                    else_=col(ApiKey.balance),
+                )
 
                 finalize_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
+                    .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance)
-                        - deducted_max_cost,
-                        balance=col(ApiKey.balance) - total_cost_msats,
-                        total_spent=col(ApiKey.total_spent) + total_cost_msats,
+                        reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                        balance=col(ApiKey.balance) - chargeable,
+                        total_spent=col(ApiKey.total_spent) + chargeable,
                     )
                 )
                 result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
 
-                # Also update total_spent and reserved_balance on the child key if it's different
                 if billing_key.hashed_key != key.hashed_key:
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
+                        .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                         .values(
-                            total_spent=col(ApiKey.total_spent) + total_cost_msats,
-                            reserved_balance=col(ApiKey.reserved_balance)
-                            - deducted_max_cost,
+                            reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                            total_spent=col(ApiKey.total_spent) + min(billing_key.balance, total_cost_msats),
                         )
                     )
                     await session.exec(child_stmt)  # type: ignore[call-overload]
@@ -889,11 +991,10 @@ async def adjust_payment_for_tokens(
                 await session.commit()
 
                 if result.rowcount:
-                    cost.total_msats = total_cost_msats
                     await session.refresh(billing_key)
                     if billing_key.hashed_key != key.hashed_key:
                         await session.refresh(key)
-
+                    cost.total_msats = total_cost_msats
                     logger.info(
                         "Finalized payment with additional charge",
                         extra={
@@ -904,9 +1005,28 @@ async def adjust_payment_for_tokens(
                             "model": model,
                         },
                     )
+                    payments_logger.info(
+                        "FINALIZE",
+                        extra={
+                            "event": "finalize",
+                            "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                            "model": model,
+                            "cost_reserved": deducted_max_cost,
+                            "cost_charged": total_cost_msats,
+                            "input_tokens": cost.input_tokens,
+                            "output_tokens": cost.output_tokens,
+                            "balance": billing_key.balance,
+                            "reserved_balance": billing_key.reserved_balance,
+                            "total_spent": billing_key.total_spent,
+                            "finalize_type": "overrun",
+                        },
+                    )
                 else:
+                    # Guard fired: reservation was already released by a concurrent
+                    # finalization for this key. Nothing left to do.
                     logger.warning(
-                        "Failed to finalize additional charge - releasing reservation",
+                        "Finalization skipped - reservation already released",
                         extra={
                             "key_hash": key.hashed_key[:8] + "...",
                             "billing_key_hash": billing_key.hashed_key[:8] + "...",
@@ -914,7 +1034,6 @@ async def adjust_payment_for_tokens(
                             "model": model,
                         },
                     )
-                    await release_reservation_only()
             else:
                 # Refund some of the base cost
                 refund = abs(cost_difference)
@@ -929,12 +1048,33 @@ async def adjust_payment_for_tokens(
                     },
                 )
 
+                if billing_key.reserved_balance < deducted_max_cost:
+                    logger.error(
+                        "reserved_balance below deducted_max_cost on refund finalization — clamping to 0",
+                        extra={
+                            "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                            "reserved_balance": billing_key.reserved_balance,
+                            "deducted_max_cost": deducted_max_cost,
+                            "total_cost_msats": total_cost_msats,
+                            "refund_amount": refund,
+                            "balance": billing_key.balance,
+                            "total_spent": billing_key.total_spent,
+                            "model": model,
+                        },
+                    )
+
+                refund_safe_reserved = case(
+                    (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                     col(ApiKey.reserved_balance) - deducted_max_cost),
+                    else_=0,
+                )
+
                 refund_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance)
-                        - deducted_max_cost,
+                        reserved_balance=refund_safe_reserved,
                         balance=col(ApiKey.balance) - total_cost_msats,
                         total_spent=col(ApiKey.total_spent) + total_cost_msats,
                     )
@@ -943,13 +1083,17 @@ async def adjust_payment_for_tokens(
 
                 # Also update total_spent and reserved_balance on the child key if it's different
                 if billing_key.hashed_key != key.hashed_key:
+                    child_refund_safe_reserved = case(
+                        (col(ApiKey.reserved_balance) >= deducted_max_cost,
+                         col(ApiKey.reserved_balance) - deducted_max_cost),
+                        else_=0,
+                    )
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
                         .values(
                             total_spent=col(ApiKey.total_spent) + total_cost_msats,
-                            reserved_balance=col(ApiKey.reserved_balance)
-                            - deducted_max_cost,
+                            reserved_balance=child_refund_safe_reserved,
                         )
                     )
                     await session.exec(child_stmt)  # type: ignore[call-overload]
@@ -986,6 +1130,24 @@ async def adjust_payment_for_tokens(
                             "model": model,
                         },
                     )
+                    payments_logger.info(
+                        "FINALIZE",
+                        extra={
+                            "event": "finalize",
+                            "key_hash": key.hashed_key[:8] + "...",
+                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                            "model": model,
+                            "cost_reserved": deducted_max_cost,
+                            "cost_charged": total_cost_msats,
+                            "refunded": refund,
+                            "input_tokens": cost.input_tokens,
+                            "output_tokens": cost.output_tokens,
+                            "balance": billing_key.balance,
+                            "reserved_balance": billing_key.reserved_balance,
+                            "total_spent": billing_key.total_spent,
+                            "finalize_type": "refund",
+                        },
+                    )
 
             return cost.dict()
 

routstr/core/admin.py

@@ -295,6 +295,7 @@ class ModelCreate(BaseModel):
     canonical_slug: str | None = None
     alias_ids: list[str] | None = None
     enabled: bool = True
+    forwarded_model_id: str | None = None
 
 
 @admin_router.post(
@@ -339,6 +340,7 @@ async def upsert_provider_model(
                 json.dumps(payload.alias_ids) if payload.alias_ids else None
             )
             existing_row.enabled = payload.enabled
+            existing_row.forwarded_model_id = payload.forwarded_model_id or payload.id
 
             session.add(existing_row)
             await session.commit()
@@ -371,6 +373,7 @@ async def upsert_provider_model(
                 ),
                 upstream_provider_id=provider_id,
                 enabled=payload.enabled,
+                forwarded_model_id=payload.forwarded_model_id or payload.id,
             )
             session.add(row)
             await session.commit()

routstr/core/db.py

@@ -105,6 +105,10 @@ class ModelRow(SQLModel, table=True):  # type: ignore
         default=None, description="JSON array of model alias IDs"
     )
     enabled: bool = Field(default=True, description="Whether this model is enabled")
+    forwarded_model_id: str | None = Field(
+        default=None,
+        description="Model ID to use when forwarding requests to upstream provider. Defaults to id if not set.",
+    )
     upstream_provider: "UpstreamProviderRow" = Relationship(back_populates="models")
 
 

routstr/core/middleware.py

@@ -38,11 +38,6 @@ class LoggingMiddleware(BaseHTTPMiddleware):
             except Exception:
                 pass
 
-        # Extract request info
-        client_host = None
-        if request.client:
-            client_host = request.client.host
-
         # Log incoming request
         logger.info(
             "Incoming request",
@@ -51,7 +46,6 @@ class LoggingMiddleware(BaseHTTPMiddleware):
                 "method": request.method,
                 "path": request.url.path,
                 "query_params": dict(request.query_params),
-                "client_host": client_host,
                 "headers": {
                     k: v
                     for k, v in request.headers.items()
@@ -100,7 +94,6 @@ class LoggingMiddleware(BaseHTTPMiddleware):
                     "path": request.url.path,
                     "status_code": response.status_code,
                     "duration_ms": round(duration * 1000, 2),
-                    "client_host": client_host,
                 },
             )
             if hasattr(response, "headers"):
@@ -120,7 +113,6 @@ class LoggingMiddleware(BaseHTTPMiddleware):
                     "method": request.method,
                     "path": request.url.path,
                     "duration_ms": round(duration * 1000, 2),
-                    "client_host": client_host,
                     "error": str(e),
                     "error_type": type(e).__name__,
                 },

routstr/core/settings.py

@@ -74,7 +74,7 @@ class Settings(BaseSettings):
     enable_pricing_refresh: bool = Field(default=True, env="ENABLE_PRICING_REFRESH")
     enable_models_refresh: bool = Field(default=True, env="ENABLE_MODELS_REFRESH")
     refund_cache_ttl_seconds: int = Field(default=3600, env="REFUND_CACHE_TTL_SECONDS")
-    refund_sweep_ttl_seconds: int = Field(default=86400, env="REFUND_SWEEP_TTL_SECONDS")
+    refund_sweep_ttl_seconds: int = Field(default=604800, env="REFUND_SWEEP_TTL_SECONDS")
 
     # Logging
     log_level: str = Field(default="INFO", env="LOG_LEVEL")

routstr/payment/models.py

@@ -60,6 +60,7 @@ class Model(BaseModel):
     upstream_provider_id: int | str | None = None
     canonical_slug: str | None = None
     alias_ids: list[str] | None = None
+    forwarded_model_id: str | None = None
 
     def __hash__(self) -> int:
         return hash(self.id)
@@ -177,14 +178,20 @@ def _row_to_model(
         upstream_provider_id=row.upstream_provider_id,
         canonical_slug=getattr(row, "canonical_slug", None),
         alias_ids=json.loads(row.alias_ids) if row.alias_ids else None,
+        forwarded_model_id=getattr(row, "forwarded_model_id", None) or row.id,
     )
 
     if apply_provider_fee:
-        (
-            parsed_pricing.max_prompt_cost,
-            parsed_pricing.max_completion_cost,
-            parsed_pricing.max_cost,
-        ) = _calculate_usd_max_costs(model)
+        max_prompt, max_completion, max_cost = _calculate_usd_max_costs(model)
+        parsed_pricing = Pricing(
+            **{
+                **parsed_pricing.dict(),
+                "max_prompt_cost": max_prompt,
+                "max_completion_cost": max_completion,
+                "max_cost": max_cost,
+            }
+        )
+        model = Model(**{**model.dict(), "pricing": parsed_pricing, "sats_pricing": None})
 
     try:
         sats_to_usd = sats_usd_price()
@@ -329,6 +336,7 @@ def _update_model_sats_pricing(model: Model, sats_to_usd: float) -> Model:
             upstream_provider_id=model.upstream_provider_id,
             canonical_slug=model.canonical_slug,
             alias_ids=model.alias_ids,
+            forwarded_model_id=model.forwarded_model_id,
         )
     except Exception as e:
         logger.error(
@@ -411,4 +419,10 @@ async def models(session: AsyncSession = Depends(get_session)) -> dict:
     from ..proxy import get_unique_models
 
     items = get_unique_models()
-    return {"data": items}
+    data = []
+    for model in items:
+        m = model.dict()
+        if model.forwarded_model_id:
+            m["id"] = model.forwarded_model_id
+        data.append(m)
+    return {"data": data}

routstr/proxy.py

@@ -207,7 +207,7 @@ async def proxy(
 
     elif auth := headers.get("authorization", None):
         key = await get_bearer_token_key(
-            headers, path, session, auth, max_cost_for_model
+            headers, path, session, auth, max_cost_for_model, model_id
         )
 
     else:
@@ -387,7 +387,12 @@ async def proxy(
 
 
 async def get_bearer_token_key(
-    headers: dict, path: str, session: AsyncSession, auth: str, min_cost: int = 0
+    headers: dict,
+    path: str,
+    session: AsyncSession,
+    auth: str,
+    min_cost: int = 0,
+    model_id: str = "unknown",
 ) -> ApiKey:
     """Handle bearer token authentication proxy requests."""
     parts = auth.split()
@@ -457,11 +462,13 @@ async def get_bearer_token_key(
     except Exception as e:
         key_preview = bearer_key[:20] + "..." if len(bearer_key) > 20 else bearer_key
         logger.error(
-            f"Bearer token validation failed: {type(e).__name__}: {e} path={path} key={key_preview!r}",
+            f"Bearer token validation failed: {type(e).__name__}: {e} path={path} model={model_id!r} min_cost={min_cost} key={key_preview!r}",
             extra={
                 "error": str(e),
                 "error_type": type(e).__name__,
                 "path": path,
+                "model_id": model_id,
+                "min_cost_msat": min_cost,
                 "bearer_key_preview": key_preview,
             },
         )

routstr/upstream/base.py

@@ -661,6 +661,9 @@ class BaseUpstreamProvider:
                 },
             )
 
+            if requested_model:
+                response_json["model"] = requested_model
+
             cost_data = await adjust_payment_for_tokens(
                 key, response_json, session, deducted_max_cost
             )
@@ -981,6 +984,9 @@ class BaseUpstreamProvider:
                 },
             )
 
+            if requested_model:
+                response_json["model"] = requested_model
+
             cost_data = await adjust_payment_for_tokens(
                 key, response_json, session, deducted_max_cost
             )
@@ -1303,7 +1309,7 @@ class BaseUpstreamProvider:
         path = self.normalize_request_path(path, model_obj)
         url = self.build_request_url(path, model_obj)
 
-        original_model_id = model_obj.id if model_obj else None
+        original_model_id = (model_obj.forwarded_model_id or model_obj.id) if model_obj else None
 
         transformed_body = self.prepare_request_body(request_body, model_obj)
 
@@ -1589,7 +1595,7 @@ class BaseUpstreamProvider:
         path = self.normalize_request_path(path, model_obj)
         url = self.build_request_url(path, model_obj)
 
-        original_model_id = model_obj.id if model_obj else None
+        original_model_id = (model_obj.forwarded_model_id or model_obj.id) if model_obj else None
 
         transformed_body = self.prepare_responses_request_body(request_body, model_obj)
 
@@ -2152,6 +2158,17 @@ class BaseUpstreamProvider:
                     },
                 )
 
+        if cost_data:
+            for i, line in enumerate(lines):
+                if line.startswith("data: "):
+                    try:
+                        data_json = json.loads(line[6:])
+                        if "usage" in data_json and data_json["usage"]:
+                            data_json["usage"]["cost_sats"] = cost_data.total_msats // 1000
+                            lines[i] = "data: " + json.dumps(data_json)
+                    except json.JSONDecodeError:
+                        pass
+
         async def generate() -> AsyncGenerator[bytes, None]:
             for line in lines:
                 yield (line + "\n").encode("utf-8")
@@ -2196,6 +2213,9 @@ class BaseUpstreamProvider:
             response_json = json.loads(content_str)
             cost_data = await self.get_x_cashu_cost(response_json, max_cost_for_model)
 
+            if cost_data and "usage" in response_json:
+                response_json["usage"]["cost_sats"] = cost_data.total_msats // 1000
+
             if not cost_data:
                 logger.error(
                     "Failed to calculate cost for response",
@@ -2262,7 +2282,7 @@ class BaseUpstreamProvider:
                 )
 
             return Response(
-                content=content_str,
+                content=json.dumps(response_json),
                 status_code=response.status_code,
                 headers=response_headers,
                 media_type="application/json",
@@ -3067,6 +3087,17 @@ class BaseUpstreamProvider:
                     },
                 )
 
+        if cost_data:
+            for i, line in enumerate(lines):
+                if line.startswith("data: "):
+                    try:
+                        data_json = json.loads(line[6:])
+                        if "usage" in data_json and data_json["usage"]:
+                            data_json["usage"]["cost_sats"] = cost_data.total_msats // 1000
+                            lines[i] = "data: " + json.dumps(data_json)
+                    except json.JSONDecodeError:
+                        pass
+
         async def generate() -> AsyncGenerator[bytes, None]:
             for line in lines:
                 yield (line + "\\n").encode("utf-8")
@@ -3099,6 +3130,9 @@ class BaseUpstreamProvider:
             response_json = json.loads(content_str)
             cost_data = await self.get_x_cashu_cost(response_json, max_cost_for_model)
 
+            if cost_data and "usage" in response_json:
+                response_json["usage"]["cost_sats"] = cost_data.total_msats // 1000
+
             if not cost_data:
                 logger.error(
                     "Failed to calculate cost for Responses API response",
@@ -3165,7 +3199,7 @@ class BaseUpstreamProvider:
                 )
 
             return Response(
-                content=content_str,
+                content=json.dumps(response_json),
                 status_code=response.status_code,
                 headers=response_headers,
                 media_type="application/json",
@@ -3356,6 +3390,7 @@ class BaseUpstreamProvider:
             upstream_provider_id=model.upstream_provider_id,
             canonical_slug=model.canonical_slug,
             alias_ids=model.alias_ids,
+            forwarded_model_id=model.forwarded_model_id,
         )
 
         (
@@ -3379,6 +3414,7 @@ class BaseUpstreamProvider:
             upstream_provider_id=model.upstream_provider_id,
             canonical_slug=model.canonical_slug,
             alias_ids=model.alias_ids,
+            forwarded_model_id=model.forwarded_model_id,
         )
 
     async def fetch_models(self) -> list[Model]:

tests/integration/conftest.py

@@ -380,6 +380,13 @@ async def integration_session(
         yield session
 
 
+@pytest_asyncio.fixture
+async def patched_db_engine(integration_engine: Any) -> AsyncGenerator[None, None]:
+    """Patch the global db engine so create_session() uses the test engine."""
+    with patch("routstr.core.db.engine", integration_engine):
+        yield
+
+
 class DatabaseSnapshot:
     """Utility to capture and compare database states"""
 

tests/integration/test_balance_negative_on_cost_overrun.py

@@ -0,0 +1,369 @@
+"""
+Integration tests for the balance-goes-negative bug in adjust_payment_for_tokens.
+
+Root cause: when actual token cost exceeds the discounted reservation
+(cost_difference > 0, caused by tolerance_percentage discounting the reservation),
+the finalization UPDATE had no WHERE guard on balance, allowing balance to go negative.
+
+Fix: added `.where(col(ApiKey.balance) >= total_cost_msats)` so the UPDATE is a no-op
+when balance is insufficient, then falls back to charging only deducted_max_cost.
+"""
+
+import uuid
+from unittest.mock import patch
+
+import pytest
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+from routstr.payment.cost_calculation import CostData
+
+
+def _make_key(balance: int, reserved: int) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"test_{uuid.uuid4().hex}",
+        balance=balance,
+        reserved_balance=reserved,
+        total_spent=0,
+        total_requests=1,
+    )
+
+
+async def _refresh(session: AsyncSession, key: ApiKey) -> ApiKey:
+    await session.refresh(key)
+    return key
+
+
+# ---------------------------------------------------------------------------
+# Helper: build a CostData where token cost > deducted_max_cost
+# ---------------------------------------------------------------------------
+
+def _cost_data(total_msats: int) -> CostData:
+    return CostData(
+        base_msats=0,
+        input_msats=total_msats // 2,
+        output_msats=total_msats - total_msats // 2,
+        total_msats=total_msats,
+        total_usd=0.0,
+        input_tokens=100,
+        output_tokens=100,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Test 1 — exact reproduction of the bug
+#
+# Setup: balance == deducted_max_cost (user has just enough for the reservation,
+#         nothing extra). Actual token cost is 1% higher (tolerance_percentage).
+#
+# Before fix: balance -= total_cost_msats → goes negative.
+# After fix:  WHERE balance >= total_cost_msats fails → fallback charges
+#             deducted_max_cost → balance reaches 0, never negative.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_balance_never_negative_when_cost_exceeds_reservation(
+    integration_session: AsyncSession,
+) -> None:
+    """Balance must not go negative when actual token cost > discounted reservation."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 990  # reserved (1% below true max of 1000)
+    actual_token_cost = 1000  # actual cost at true max
+
+    # User has balance exactly equal to the reservation — tight budget
+    key = _make_key(balance=deducted_max_cost, reserved=deducted_max_cost)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {"model": "test-model", "usage": {"prompt_tokens": 100, "completion_tokens": 100}}
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(key, response_data, integration_session, deducted_max_cost)
+
+    await _refresh(integration_session, key)
+
+    assert key.balance >= 0, f"Balance went negative: {key.balance}"
+    assert key.reserved_balance >= 0, f"Reserved balance went negative: {key.reserved_balance}"
+    assert key.reserved_balance == 0, "Reservation must be fully released after finalization"
+
+
+# ---------------------------------------------------------------------------
+# Test 2 — balance is ZERO after the reservation is accounted for
+#           (absolute floor case)
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_balance_floor_at_zero_on_overrun(
+    integration_session: AsyncSession,
+) -> None:
+    """When balance exactly covers deducted_max_cost and cost overruns, balance reaches 0 not negative."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 500
+    actual_token_cost = 550  # 10% overrun
+
+    key = _make_key(balance=500, reserved=500)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {"model": "test-model", "usage": {"prompt_tokens": 50, "completion_tokens": 50}}
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(key, response_data, integration_session, deducted_max_cost)
+
+    await _refresh(integration_session, key)
+
+    assert key.balance == 0, (
+        f"Expected balance=0 (charged deducted_max_cost fallback), got {key.balance}"
+    )
+    assert key.reserved_balance == 0, f"Reserved balance should be 0, got {key.reserved_balance}"
+    # Fallback charges deducted_max_cost
+    assert key.total_spent == deducted_max_cost, (
+        f"Expected total_spent={deducted_max_cost}, got {key.total_spent}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Test 3 — balance has enough room: full token cost should be charged
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_full_cost_charged_when_balance_sufficient_for_overrun(
+    integration_session: AsyncSession,
+) -> None:
+    """When balance covers total_cost_msats, the full amount is charged (not just deducted_max_cost)."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 990
+    actual_token_cost = 1000
+
+    # User has extra balance beyond the reservation
+    key = _make_key(balance=2000, reserved=990)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {"model": "test-model", "usage": {"prompt_tokens": 100, "completion_tokens": 100}}
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(key, response_data, integration_session, deducted_max_cost)
+
+    await _refresh(integration_session, key)
+
+    assert key.balance >= 0, f"Balance went negative: {key.balance}"
+    assert key.reserved_balance == 0, f"Reservation not released: {key.reserved_balance}"
+    assert key.total_spent == actual_token_cost, (
+        f"Expected full charge of {actual_token_cost}, got {key.total_spent}"
+    )
+    assert key.balance == 2000 - actual_token_cost, (
+        f"Expected balance={2000 - actual_token_cost}, got {key.balance}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Test 4 — concurrent finalizations with cost overrun
+#
+# Multiple requests finish concurrently. Each has a small overrun.
+# None should drive balance negative.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_concurrent_cost_overruns_never_negative(
+    integration_session: AsyncSession,
+    patched_db_engine: None,
+) -> None:
+    """Concurrent finalization with cost overruns must never produce negative balance."""
+    import asyncio
+
+    from routstr.auth import adjust_payment_for_tokens, pay_for_request
+    from routstr.core.db import create_session
+
+    deducted_max_cost = 990
+    actual_token_cost = 1000
+    n_requests = 5
+
+    # Fund the key with exactly enough for n_requests reservations + a tiny buffer
+    starting_balance = deducted_max_cost * n_requests
+    key_hash = f"test_concurrent_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        key = ApiKey(
+            hashed_key=key_hash,
+            balance=starting_balance,
+            reserved_balance=0,
+            total_spent=0,
+            total_requests=0,
+        )
+        session.add(key)
+        await session.commit()
+
+    # Reserve n_requests slots (sequentially, as pay_for_request is atomic)
+    async with create_session() as session:
+        key_to_reserve = await session.get(ApiKey, key_hash)
+        assert key_to_reserve is not None
+        for _ in range(n_requests):
+            await pay_for_request(key_to_reserve, deducted_max_cost, session)
+            await session.refresh(key_to_reserve)
+
+    # Now finalize all concurrently with cost overrun
+    async def finalize() -> None:
+        response_data = {
+            "model": "test-model",
+            "usage": {"prompt_tokens": 100, "completion_tokens": 100},
+        }
+        async with create_session() as session:
+            fresh_key = await session.get(ApiKey, key_hash)
+            assert fresh_key is not None
+            with patch(
+                "routstr.auth.calculate_cost",
+                return_value=_cost_data(actual_token_cost),
+            ):
+                await adjust_payment_for_tokens(
+                    fresh_key, response_data, session, deducted_max_cost
+                )
+
+    await asyncio.gather(*[finalize() for _ in range(n_requests)])
+
+    async with create_session() as session:
+        final_key = await session.get(ApiKey, key_hash)
+        assert final_key is not None
+
+    assert final_key.balance >= 0, (
+        f"Balance went negative after concurrent overruns: {final_key.balance}"
+    )
+    assert final_key.reserved_balance == 0, (
+        f"Reserved balance not fully released: {final_key.reserved_balance}"
+    )
+    assert final_key.total_spent <= starting_balance, (
+        f"Total spent ({final_key.total_spent}) exceeds starting balance ({starting_balance})"
+    )
+    # Every request must have been charged at least deducted_max_cost — no free inference.
+    assert final_key.total_spent == starting_balance, (
+        f"Expected total_spent={starting_balance} (all {n_requests} reservations charged), "
+        f"got {final_key.total_spent} — at least one request got free inference"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Test 5 — overrun with no balance at all (reserved_balance == balance)
+#           simulates a user who topped up to exactly the reservation floor
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_zero_free_balance_overrun_is_safe(
+    integration_session: AsyncSession,
+) -> None:
+    """User with zero free balance (all reserved) should never go negative on overrun."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 1000
+    actual_token_cost = 1050
+
+    # balance == reserved_balance: zero free balance
+    key = _make_key(balance=1000, reserved=1000)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {"model": "test-model", "usage": {"prompt_tokens": 50, "completion_tokens": 100}}
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(key, response_data, integration_session, deducted_max_cost)
+
+    await _refresh(integration_session, key)
+
+    assert key.balance >= 0, f"Balance went negative: {key.balance}"
+    assert key.reserved_balance >= 0, f"Reserved balance went negative: {key.reserved_balance}"
+
+
+# ---------------------------------------------------------------------------
+# Test 6 — parallel requests: second finalization must not get free inference
+#
+# Root cause of the bug fixed in auth.py:
+#   `.where(col(ApiKey.balance) >= total_cost_msats)` ignores other requests'
+#   reservations, so after Request A charges total_cost_msats, balance can drop
+#   below deducted_max_cost, causing Request B's fallback to release for free.
+#
+# Fix: use `balance - reserved_balance + deducted_max_cost >= total_cost_msats`
+#   so the check accounts for concurrent reservations.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_parallel_requests_no_free_inference(
+    integration_session: AsyncSession,
+    patched_db_engine: None,
+) -> None:
+    """Second parallel finalization must be charged even when first depleted free balance."""
+    import asyncio
+
+    from routstr.auth import adjust_payment_for_tokens
+    from routstr.core.db import create_session
+
+    deducted_max_cost = 100
+    actual_token_cost = 150  # overrun: 50 more than reserved
+
+    # Fund the key with exactly 2 * deducted_max_cost.
+    # Both requests pre-reserved 100 each → balance=200, reserved=200, free=0.
+    # Old check (balance >= total_cost_msats):
+    #   Request A: 200 >= 150 ✓ → charges 150 → balance=50, reserved=100
+    #   Request B: 50 >= 150 ✗ → fallback: 50 >= 100 ✗ → releases FREE
+    # New check (balance - reserved + deducted >= total_cost_msats):
+    #   Both fall to fallback (0 free balance).
+    #   Both charge deducted_max_cost=100 → total_spent=200, balance=0.
+    starting_balance = deducted_max_cost * 2
+    key_hash = f"test_parallel_no_free_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        key = ApiKey(
+            hashed_key=key_hash,
+            balance=starting_balance,
+            reserved_balance=deducted_max_cost * 2,  # both slots pre-reserved
+            total_spent=0,
+            total_requests=2,
+        )
+        session.add(key)
+        await session.commit()
+
+    async def finalize() -> None:
+        response_data = {
+            "model": "test-model",
+            "usage": {"prompt_tokens": 50, "completion_tokens": 100},
+        }
+        async with create_session() as session:
+            fresh_key = await session.get(ApiKey, key_hash)
+            assert fresh_key is not None
+            with patch(
+                "routstr.auth.calculate_cost",
+                return_value=_cost_data(actual_token_cost),
+            ):
+                await adjust_payment_for_tokens(
+                    fresh_key, response_data, session, deducted_max_cost
+                )
+
+    await asyncio.gather(finalize(), finalize())
+
+    async with create_session() as session:
+        final_key = await session.get(ApiKey, key_hash)
+        assert final_key is not None
+
+    assert final_key.balance >= 0, f"Balance went negative: {final_key.balance}"
+    assert final_key.reserved_balance == 0, (
+        f"Reserved balance not released: {final_key.reserved_balance}"
+    )
+    # Both requests must have been charged — no free inference.
+    assert final_key.total_spent == starting_balance, (
+        f"Expected total_spent={starting_balance} (both reservations charged), "
+        f"got {final_key.total_spent} — one request got free inference"
+    )

tests/integration/test_insufficient_balance.py

@@ -0,0 +1,240 @@
+"""
+Tests showing how a user hits "Insufficient balance: X mSats required for this model"
+when their balance is too low for the model's cost.
+
+The log line that triggered this:
+  WARNING  Insufficient billing balance during validation
+  ERROR    Bearer token validation failed: HTTPException: 402:
+           {'error': {'message': 'Insufficient balance: 622888 mSats required
+           for this model. 20320 available.', ...}}
+
+This happens in validate_bearer_key (auth.py) when:
+  billing_key.total_balance < min_cost   (model's max cost)
+
+and also in pay_for_request when the atomic UPDATE finds no available balance.
+"""
+
+import uuid
+from unittest.mock import patch
+
+import pytest
+from fastapi import HTTPException
+from httpx import AsyncClient
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+
+
+def _key(balance: int, reserved: int = 0) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"test_{uuid.uuid4().hex}",
+        balance=balance,
+        reserved_balance=reserved,
+        total_spent=0,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Test 1 — simplest case: balance < model cost → pay_for_request raises 402
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_pay_for_request_raises_402_when_balance_too_low(
+    integration_session: AsyncSession,
+) -> None:
+    """
+    User has 20_000 msats. Model costs 622_888 msats.
+    pay_for_request must raise HTTP 402 with a clear message.
+    """
+    from routstr.auth import pay_for_request
+
+    model_cost = 622_888
+    user_balance = 20_000
+
+    key = _key(balance=user_balance)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, model_cost, integration_session)
+
+    assert exc_info.value.status_code == 402
+    detail = exc_info.value.detail
+    assert isinstance(detail, dict)
+    error = detail["error"]
+    assert error["code"] == "insufficient_balance"
+    assert str(model_cost) in error["message"]
+    assert str(user_balance) in error["message"]
+
+    # Balance must be untouched
+    await integration_session.refresh(key)
+    assert key.balance == user_balance
+    assert key.reserved_balance == 0
+
+
+# ---------------------------------------------------------------------------
+# Test 2 — balance is zero
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_pay_for_request_raises_402_on_zero_balance(
+    integration_session: AsyncSession,
+) -> None:
+    """User with zero balance cannot make any request."""
+    from routstr.auth import pay_for_request
+
+    key = _key(balance=0)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, 1_000, integration_session)
+
+    assert exc_info.value.status_code == 402
+    detail = exc_info.value.detail
+    assert isinstance(detail, dict)
+    assert detail["error"]["code"] == "insufficient_balance"
+
+
+# ---------------------------------------------------------------------------
+# Test 3 — all balance is reserved (total_balance = balance - reserved = 0)
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_pay_for_request_raises_402_when_all_balance_reserved(
+    integration_session: AsyncSession,
+) -> None:
+    """
+    User has 50_000 msats balance but 50_000 is already reserved for in-flight
+    requests. Free balance (total_balance) = 0. Should get 402.
+    """
+    from routstr.auth import pay_for_request
+
+    key = _key(balance=50_000, reserved=50_000)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, 1_000, integration_session)
+
+    assert exc_info.value.status_code == 402
+    # Balance and reserved must be untouched
+    await integration_session.refresh(key)
+    assert key.balance == 50_000
+    assert key.reserved_balance == 50_000
+
+
+# ---------------------------------------------------------------------------
+# Test 4 — balance just one msat below model cost
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_pay_for_request_raises_402_one_msat_short(
+    integration_session: AsyncSession,
+) -> None:
+    """Off-by-one: balance is exactly model_cost - 1."""
+    from routstr.auth import pay_for_request
+
+    model_cost = 10_000
+    key = _key(balance=model_cost - 1)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, model_cost, integration_session)
+
+    assert exc_info.value.status_code == 402
+    await integration_session.refresh(key)
+    assert key.balance == model_cost - 1   # untouched
+    assert key.reserved_balance == 0
+
+
+# ---------------------------------------------------------------------------
+# Test 5 — balance exactly equal to model cost → succeeds
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_pay_for_request_succeeds_when_balance_equals_cost(
+    integration_session: AsyncSession,
+) -> None:
+    """Balance == model cost: the request should be reserved successfully."""
+    from routstr.auth import pay_for_request
+
+    model_cost = 10_000
+    key = _key(balance=model_cost)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    # Should not raise
+    await pay_for_request(key, model_cost, integration_session)
+
+    await integration_session.refresh(key)
+    assert key.reserved_balance == model_cost
+    assert key.balance == model_cost   # balance unchanged, only reserved goes up
+
+
+# ---------------------------------------------------------------------------
+# Test 6 — HTTP layer returns 402 JSON with the right shape
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_http_402_response_shape_on_insufficient_balance(
+    integration_client: AsyncClient,
+    integration_session: AsyncSession,
+) -> None:
+    """
+    End-to-end: POST /v1/chat/completions with a key whose balance is far below
+    the mocked model cost returns HTTP 402 with the expected JSON error body.
+
+    Matches exactly the log snippet in the bug report:
+      'Insufficient balance: X mSats required for this model. Y available.'
+    """
+    from unittest.mock import AsyncMock, MagicMock
+
+    model_cost = 622_888
+    user_balance = 20_320
+
+    key = _key(balance=user_balance)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    # Minimal model stub so proxy routing doesn't 400 before reaching balance check
+    mock_model = MagicMock()
+    mock_model.sats_pricing = None
+
+    # Upstream stub — never reached because balance check fires first
+    mock_upstream = MagicMock()
+    mock_upstream.prepare_headers = MagicMock(return_value={})
+
+    with (
+        patch("routstr.proxy.get_model_instance", return_value=mock_model),
+        patch("routstr.proxy.get_provider_for_model", return_value=[mock_upstream]),
+        # Patch where it is used (proxy imports it at module level)
+        patch(
+            "routstr.proxy.get_max_cost_for_model",
+            new=AsyncMock(return_value=model_cost),
+        ),
+    ):
+        response = await integration_client.post(
+            "/v1/chat/completions",
+            headers={"Authorization": f"Bearer sk-{key.hashed_key}"},
+            json={
+                "model": "gpt-4o",
+                "messages": [{"role": "user", "content": "hello"}],
+            },
+        )
+
+    assert response.status_code == 402
+    body = response.json()
+    # FastAPI wraps HTTPException detail under "detail"
+    error = body["detail"]["error"]
+    assert error["code"] == "insufficient_balance"
+    assert error["type"] == "insufficient_quota"
+    assert str(model_cost) in error["message"]
+    assert str(user_balance) in error["message"]
+
+    # Balance must be completely untouched
+    await integration_session.refresh(key)
+    assert key.balance == user_balance
+    assert key.reserved_balance == 0
+    assert key.total_spent == 0

tests/integration/test_reservation_lifecycle.py

@@ -0,0 +1,268 @@
+"""
+Tests for the reservation lifecycle:
+
+  1. Reserve  → reserved_balance increases, available (total_balance) decreases.
+  2. Reserve  → revert  → reserved_balance restored, balance untouched.
+  3. Reserve  → finalise → reserved_balance released, balance charged.
+  4. Two parallel reserves, only one fits → second blocked with 402.
+  5. Three parallel reserves, two fit, third blocked with 402.
+  6. Sequential reserves until balance exhausted → next request blocked.
+
+Reservation invariant enforced by the atomic WHERE clause in pay_for_request:
+    balance - reserved_balance >= cost_per_request
+"""
+
+import asyncio
+import uuid
+
+import pytest
+from fastapi import HTTPException
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.auth import pay_for_request, revert_pay_for_request
+from routstr.core.db import ApiKey, create_session
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _make_key(balance: int, reserved: int = 0) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"test_{uuid.uuid4().hex}",
+        balance=balance,
+        reserved_balance=reserved,
+        total_spent=0,
+        total_requests=0,
+    )
+
+
+async def _persist(session: AsyncSession, key: ApiKey) -> ApiKey:
+    session.add(key)
+    await session.commit()
+    await session.refresh(key)
+    return key
+
+
+# ---------------------------------------------------------------------------
+# Test 1 — Reserve: reserved_balance increases, available balance decreases
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_reserve_increases_reserved_balance(
+    integration_session: AsyncSession,
+) -> None:
+    """pay_for_request must increment reserved_balance by cost_per_request."""
+    cost = 100
+    key = await _persist(integration_session, _make_key(balance=500))
+
+    await pay_for_request(key, cost, integration_session)
+    await integration_session.refresh(key)
+
+    assert key.reserved_balance == cost
+    assert key.balance == 500  # balance column is NOT decremented on reserve
+    assert key.total_balance == 500 - cost  # available = balance - reserved
+
+
+# ---------------------------------------------------------------------------
+# Test 2 — Revert: reserved_balance restored, balance untouched
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_revert_releases_reservation(
+    integration_session: AsyncSession,
+) -> None:
+    """revert_pay_for_request must release the reservation without touching balance."""
+    cost = 150
+    key = await _persist(integration_session, _make_key(balance=300))
+
+    await pay_for_request(key, cost, integration_session)
+    await integration_session.refresh(key)
+    assert key.reserved_balance == cost
+
+    await revert_pay_for_request(key, integration_session, cost)
+    await integration_session.refresh(key)
+
+    assert key.reserved_balance == 0
+    assert key.balance == 300  # balance unchanged after revert
+
+
+# ---------------------------------------------------------------------------
+# Test 3 — Finalise: reservation released + balance charged
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_finalise_releases_reservation_and_charges_balance(
+    integration_session: AsyncSession,
+) -> None:
+    """adjust_payment_for_tokens must zero reserved_balance and deduct actual cost."""
+    from unittest.mock import patch
+
+    from routstr.auth import adjust_payment_for_tokens
+    from routstr.payment.cost_calculation import CostData
+
+    cost = 100
+    actual = 80  # actual < reserved → refund path
+    key = await _persist(integration_session, _make_key(balance=500))
+
+    await pay_for_request(key, cost, integration_session)
+    await integration_session.refresh(key)
+    assert key.reserved_balance == cost
+
+    cost_data = CostData(
+        base_msats=0,
+        input_msats=40,
+        output_msats=40,
+        total_msats=actual,
+        total_usd=0.0,
+        input_tokens=50,
+        output_tokens=50,
+    )
+    response_data = {"model": "test-model", "usage": {"prompt_tokens": 50, "completion_tokens": 50}}
+
+    with patch("routstr.auth.calculate_cost", return_value=cost_data):
+        await adjust_payment_for_tokens(key, response_data, integration_session, cost)
+
+    await integration_session.refresh(key)
+
+    assert key.reserved_balance == 0
+    assert key.balance == 500 - actual
+    assert key.total_spent == actual
+
+
+# ---------------------------------------------------------------------------
+# Test 4 — Concurrent: second parallel reserve blocked when balance exhausted
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_concurrent_second_reserve_blocked_when_balance_exhausted(
+    patched_db_engine: None,
+) -> None:
+    """When two requests race for the same balance, only one succeeds; the other gets 402."""
+    cost = 300
+    key_hash = f"test_concurrent_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        key = ApiKey(
+            hashed_key=key_hash,
+            balance=300,       # exactly enough for ONE reservation
+            reserved_balance=0,
+            total_spent=0,
+            total_requests=0,
+        )
+        session.add(key)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt_reserve() -> None:
+        async with create_session() as session:
+            fresh_key = await session.get(ApiKey, key_hash)
+            assert fresh_key is not None
+            try:
+                await pay_for_request(fresh_key, cost, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(attempt_reserve(), attempt_reserve())
+
+    assert sorted(results) == ["blocked", "success"], (
+        f"Expected exactly one success and one 402, got: {results}"
+    )
+
+    async with create_session() as session:
+        final = await session.get(ApiKey, key_hash)
+        assert final is not None
+
+    # reserved_balance must equal exactly one reservation (not two)
+    assert final.reserved_balance == cost, (
+        f"Expected reserved_balance={cost}, got {final.reserved_balance}"
+    )
+    assert final.balance == 300, "Balance column must not be modified by reservation"
+
+
+# ---------------------------------------------------------------------------
+# Test 5 — Concurrent: three requests, two fit, third blocked
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_three_parallel_reserves_third_blocked(
+    patched_db_engine: None,
+) -> None:
+    """Balance covers two reservations exactly; the third concurrent request must be blocked."""
+    cost = 100
+    key_hash = f"test_three_parallel_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        key = ApiKey(
+            hashed_key=key_hash,
+            balance=200,       # fits exactly 2 reservations of 100
+            reserved_balance=0,
+            total_spent=0,
+            total_requests=0,
+        )
+        session.add(key)
+        await session.commit()
+
+    results: list[str] = []
+
+    async def attempt_reserve() -> None:
+        async with create_session() as session:
+            fresh_key = await session.get(ApiKey, key_hash)
+            assert fresh_key is not None
+            try:
+                await pay_for_request(fresh_key, cost, session)
+                results.append("success")
+            except HTTPException as exc:
+                assert exc.status_code == 402
+                results.append("blocked")
+
+    await asyncio.gather(
+        attempt_reserve(),
+        attempt_reserve(),
+        attempt_reserve(),
+    )
+
+    successes = results.count("success")
+    blocked = results.count("blocked")
+
+    assert successes == 2, f"Expected 2 successes, got {successes}: {results}"
+    assert blocked == 1, f"Expected 1 blocked, got {blocked}: {results}"
+
+    async with create_session() as session:
+        final = await session.get(ApiKey, key_hash)
+        assert final is not None
+
+    assert final.reserved_balance == cost * 2, (
+        f"Expected reserved_balance={cost * 2}, got {final.reserved_balance}"
+    )
+    assert final.balance == 200, "Balance column must not be modified by reservation"
+
+
+# ---------------------------------------------------------------------------
+# Test 6 — Sequential exhaustion: reserve until empty, next request blocked
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_sequential_reserves_block_when_balance_exhausted(
+    integration_session: AsyncSession,
+) -> None:
+    """Repeated reservations should block as soon as available balance drops below cost."""
+    cost = 100
+    key = await _persist(integration_session, _make_key(balance=250))
+
+    # First two succeed (100 + 100 = 200 ≤ 250)
+    await pay_for_request(key, cost, integration_session)
+    await pay_for_request(key, cost, integration_session)
+    await integration_session.refresh(key)
+    assert key.reserved_balance == 200
+    assert key.total_balance == 50  # 250 - 200
+
+    # Third: only 50 available, need 100 → blocked
+    with pytest.raises(HTTPException) as exc_info:
+        await pay_for_request(key, cost, integration_session)
+
+    assert exc_info.value.status_code == 402
+    await integration_session.refresh(key)
+    assert key.reserved_balance == 200  # unchanged after failed reserve

tests/unit/test_x_cashu_cost_sats.py

@@ -0,0 +1,216 @@
+import json
+import os
+from unittest.mock import AsyncMock, patch
+
+import httpx
+import pytest
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+
+from routstr.payment.cost_calculation import CostData  # noqa: E402
+from routstr.upstream.base import BaseUpstreamProvider  # noqa: E402
+
+
+def _make_provider() -> BaseUpstreamProvider:
+    return BaseUpstreamProvider(base_url="http://test", api_key="test-key")
+
+
+def _make_httpx_response(status_code: int = 200) -> httpx.Response:
+    return httpx.Response(status_code, headers={})
+
+
+def _make_cost_data(total_msats: int = 5000) -> CostData:
+    return CostData(
+        base_msats=0,
+        input_msats=3000,
+        output_msats=2000,
+        total_msats=total_msats,
+        total_usd=0.00025,
+        input_tokens=100,
+        output_tokens=50,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Non-streaming (chat completions)
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_non_streaming_includes_cost_sats() -> None:
+    provider = _make_provider()
+    cost_data = _make_cost_data(total_msats=5000)
+
+    response_body = {
+        "model": "gpt-4o",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "total_tokens": 150,
+            "cost": 0.00025,
+        },
+    }
+    content_str = json.dumps(response_body)
+    httpx_response = _make_httpx_response()
+
+    with (
+        patch.object(provider, "get_x_cashu_cost", new=AsyncMock(return_value=cost_data)),
+        patch.object(provider, "send_refund", new=AsyncMock(return_value="cashuA_refund_token")),
+    ):
+        response = await provider.handle_x_cashu_non_streaming_response(
+            content_str=content_str,
+            response=httpx_response,
+            amount=10000,
+            unit="msat",
+            max_cost_for_model=10000,
+            mint=None,
+            payment_token_hash=None,
+        )
+
+    body = json.loads(response.body)
+    assert "cost_sats" in body["usage"]
+    assert body["usage"]["cost_sats"] == 5  # 5000 msats // 1000
+
+
+@pytest.mark.asyncio
+async def test_non_streaming_cost_sats_value_rounds_down() -> None:
+    provider = _make_provider()
+    cost_data = _make_cost_data(total_msats=1999)
+
+    response_body = {"model": "gpt-4o", "usage": {"prompt_tokens": 10}}
+    content_str = json.dumps(response_body)
+
+    with (
+        patch.object(provider, "get_x_cashu_cost", new=AsyncMock(return_value=cost_data)),
+        patch.object(provider, "send_refund", new=AsyncMock(return_value="cashuA_refund_token")),
+    ):
+        response = await provider.handle_x_cashu_non_streaming_response(
+            content_str=content_str,
+            response=_make_httpx_response(),
+            amount=10000,
+            unit="msat",
+            max_cost_for_model=10000,
+        )
+
+    body = json.loads(response.body)
+    assert body["usage"]["cost_sats"] == 1  # 1999 // 1000
+
+
+@pytest.mark.asyncio
+async def test_non_streaming_preserves_existing_usage_fields() -> None:
+    provider = _make_provider()
+    cost_data = _make_cost_data(total_msats=3000)
+
+    response_body = {
+        "model": "gpt-4o",
+        "usage": {
+            "prompt_tokens": 100,
+            "completion_tokens": 50,
+            "total_tokens": 150,
+            "cost": 0.00015,
+        },
+    }
+
+    with (
+        patch.object(provider, "get_x_cashu_cost", new=AsyncMock(return_value=cost_data)),
+        patch.object(provider, "send_refund", new=AsyncMock(return_value="cashuA_refund_token")),
+    ):
+        response = await provider.handle_x_cashu_non_streaming_response(
+            content_str=json.dumps(response_body),
+            response=_make_httpx_response(),
+            amount=10000,
+            unit="msat",
+            max_cost_for_model=10000,
+        )
+
+    body = json.loads(response.body)
+    usage = body["usage"]
+    assert usage["prompt_tokens"] == 100
+    assert usage["completion_tokens"] == 50
+    assert usage["total_tokens"] == 150
+    assert usage["cost"] == 0.00015
+    assert usage["cost_sats"] == 3
+
+
+# ---------------------------------------------------------------------------
+# Streaming (chat completions)
+# ---------------------------------------------------------------------------
+
+async def _collect_streaming(response: object) -> list[str]:
+    chunks: list[str] = []
+    async for chunk in response.body_iterator:  # type: ignore[attr-defined]
+        if isinstance(chunk, bytes):
+            chunks.append(chunk.decode("utf-8"))
+        else:
+            chunks.append(str(chunk))
+    return chunks
+
+
+@pytest.mark.asyncio
+async def test_streaming_includes_cost_sats_in_usage_chunk() -> None:
+    provider = _make_provider()
+    cost_data = _make_cost_data(total_msats=7000)
+
+    usage_chunk = {
+        "id": "chatcmpl-123",
+        "model": "gpt-4o",
+        "usage": {"prompt_tokens": 100, "completion_tokens": 50, "total_tokens": 150},
+    }
+    content_str = "\n".join([
+        'data: {"id":"chatcmpl-123","model":"gpt-4o","choices":[]}',
+        f"data: {json.dumps(usage_chunk)}",
+        "data: [DONE]",
+    ])
+
+    with patch.object(provider, "get_x_cashu_cost", new=AsyncMock(return_value=cost_data)):
+        response = await provider.handle_x_cashu_streaming_response(
+            content_str=content_str,
+            response=_make_httpx_response(),
+            amount=10000,
+            unit="msat",
+            max_cost_for_model=10000,
+            mint=None,
+            payment_token_hash=None,
+        )
+
+    chunks = await _collect_streaming(response)
+
+    full_output = "".join(chunks)
+    usage_line = next(
+        line for line in full_output.split("\n") if '"usage"' in line and "cost_sats" in line
+    )
+    data_json = json.loads(usage_line.lstrip("data: ").strip())
+    assert data_json["usage"]["cost_sats"] == 7  # 7000 // 1000
+
+
+@pytest.mark.asyncio
+async def test_streaming_non_usage_chunks_unmodified() -> None:
+    provider = _make_provider()
+    cost_data = _make_cost_data(total_msats=2000)
+
+    regular_chunk = {"id": "chatcmpl-123", "model": "gpt-4o", "choices": [{"delta": {"content": "hi"}}]}
+    usage_chunk = {"id": "chatcmpl-123", "model": "gpt-4o", "usage": {"prompt_tokens": 10}}
+    content_str = "\n".join([
+        f"data: {json.dumps(regular_chunk)}",
+        f"data: {json.dumps(usage_chunk)}",
+        "data: [DONE]",
+    ])
+
+    with patch.object(provider, "get_x_cashu_cost", new=AsyncMock(return_value=cost_data)):
+        response = await provider.handle_x_cashu_streaming_response(
+            content_str=content_str,
+            response=_make_httpx_response(),
+            amount=10000,
+            unit="msat",
+            max_cost_for_model=10000,
+        )
+
+    chunks = await _collect_streaming(response)
+
+    lines = [
+        line for line in "".join(chunks).split("\n")
+        if line.startswith("data: ") and line != "data: [DONE]"
+    ]
+    regular_line_data = json.loads(lines[0][6:])
+    # regular chunk should not have cost_sats injected
+    assert "cost_sats" not in regular_line_data.get("usage", {})

ui/components/add-provider-model-dialog.tsx

@@ -1,6 +1,6 @@
 'use client';
 
-import React, { useEffect, useMemo, useState } from 'react';
+import React, { useCallback, useEffect, useMemo, useState } from 'react';
 import { useForm } from 'react-hook-form';
 import { z } from 'zod';
 import { zodResolver } from '@hookform/resolvers/zod';
@@ -39,7 +39,7 @@ import {
   FormMessage,
 } from '@/components/ui/form';
 import { Switch } from '@/components/ui/switch';
-import { Loader2, Plus } from 'lucide-react';
+import { Check, Copy, Loader2, Plus } from 'lucide-react';
 import { toast } from 'sonner';
 import { AdminService, type AdminModel } from '@/lib/api/services/admin';
 
@@ -64,6 +64,7 @@ const FormSchema = z.object({
   instruct_type: z.string().default(''),
   canonical_slug: z.string().default(''),
   alias_ids_raw: z.string().default(''),
+  forwarded_model_id: z.string().default(''),
   upstream_provider_id: z.string().default(''),
   input_cost: z.coerce.number().min(0).default(0),
   output_cost: z.coerce.number().min(0).default(0),
@@ -104,6 +105,7 @@ export function AddProviderModelDialog({
   const [isPresetOpen, setIsPresetOpen] = useState(false);
   const [selectedPresetLabel, setSelectedPresetLabel] =
     useState('Select a preset');
+  const [forwardedModelIdCopied, setForwardedModelIdCopied] = useState(false);
 
   const form = useForm<FormData>({
     resolver: zodResolver(FormSchema) as never,
@@ -119,6 +121,7 @@ export function AddProviderModelDialog({
       instruct_type: '',
       canonical_slug: '',
       alias_ids_raw: '',
+      forwarded_model_id: '',
       upstream_provider_id: '',
       input_cost: 0,
       output_cost: 0,
@@ -180,6 +183,7 @@ export function AddProviderModelDialog({
             : '',
         canonical_slug: initialData.canonical_slug || '',
         alias_ids_raw: listToString(initialData.alias_ids),
+        forwarded_model_id: initialData.forwarded_model_id || initialData.id,
         upstream_provider_id:
           typeof initialData.upstream_provider_id === 'string'
             ? initialData.upstream_provider_id
@@ -223,6 +227,7 @@ export function AddProviderModelDialog({
         instruct_type: '',
         canonical_slug: '',
         alias_ids_raw: '',
+        forwarded_model_id: '',
         upstream_provider_id: '',
         input_cost: 0,
         output_cost: 0,
@@ -280,6 +285,7 @@ export function AddProviderModelDialog({
     );
     form.setValue('canonical_slug', model.canonical_slug || '');
     form.setValue('alias_ids_raw', listToString(model.alias_ids));
+    form.setValue('forwarded_model_id', model.forwarded_model_id || model.id);
     form.setValue(
       'upstream_provider_id',
       typeof model.upstream_provider_id === 'string'
@@ -385,6 +391,7 @@ export function AddProviderModelDialog({
         canonical_slug: data.canonical_slug?.trim() || null,
         alias_ids: listFromString(data.alias_ids_raw || ''),
         enabled: data.enabled,
+        forwarded_model_id: data.forwarded_model_id?.trim() || data.id,
       };
 
       if (isEdit) {
@@ -520,6 +527,53 @@ export function AddProviderModelDialog({
                 )}
               />
 
+              <FormField
+                control={form.control}
+                name='forwarded_model_id'
+                render={({ field }) => {
+                  const handleCopy = () => {
+                    const value = field.value || form.getValues('id');
+                    if (!value) return;
+                    navigator.clipboard.writeText(value);
+                    setForwardedModelIdCopied(true);
+                    setTimeout(() => setForwardedModelIdCopied(false), 1500);
+                  };
+                  return (
+                    <FormItem>
+                      <FormLabel>Upstream Model ID</FormLabel>
+                      <FormControl>
+                        <div className='flex gap-2'>
+                          <Input
+                            placeholder={
+                              form.watch('id') || 'e.g., openai/gpt-4o'
+                            }
+                            {...field}
+                          />
+                          <Button
+                            type='button'
+                            variant='outline'
+                            size='icon'
+                            onClick={handleCopy}
+                            title='Copy model ID'
+                          >
+                            {forwardedModelIdCopied ? (
+                              <Check className='h-4 w-4 text-green-500' />
+                            ) : (
+                              <Copy className='h-4 w-4' />
+                            )}
+                          </Button>
+                        </div>
+                      </FormControl>
+                      <FormDescription>
+                        Model ID sent to the upstream provider. Defaults to the
+                        model&apos;s own ID.
+                      </FormDescription>
+                      <FormMessage />
+                    </FormItem>
+                  );
+                }}
+              />
+
               <FormField
                 control={form.control}
                 name='name'

ui/components/edit-model-form.tsx

@@ -1,577 +0,0 @@
-'use client';
-
-import React, { useState, useEffect, useCallback } from 'react';
-import { useForm } from 'react-hook-form';
-import { zodResolver } from '@hookform/resolvers/zod';
-import { z } from 'zod';
-import { type Model } from '@/lib/api/schemas/models';
-import { AdminService, type AdminModel } from '@/lib/api/services/admin';
-import { Button } from '@/components/ui/button';
-import { Input } from '@/components/ui/input';
-import { Textarea } from '@/components/ui/textarea';
-import {
-  Dialog,
-  DialogContent,
-  DialogDescription,
-  DialogHeader,
-  DialogTitle,
-} from '@/components/ui/dialog';
-import {
-  Form,
-  FormControl,
-  FormDescription,
-  FormField,
-  FormItem,
-  FormLabel,
-  FormMessage,
-} from '@/components/ui/form';
-import { Edit3, Loader2 } from 'lucide-react';
-import { toast } from 'sonner';
-import { Switch } from '@/components/ui/switch';
-
-const EditModelFormSchema = z.object({
-  name: z.string().min(1, 'Name is required'),
-  description: z.string().optional(),
-  context_length: z.number().min(0),
-  prompt: z.number().min(0),
-  completion: z.number().min(0),
-  enabled: z.boolean(),
-});
-
-type EditModelFormData = z.infer<typeof EditModelFormSchema>;
-
-const roundToFiveDecimals = (value: number | undefined | null): number => {
-  if (value === undefined || value === null || isNaN(value)) {
-    return 0;
-  }
-  return Math.round(value * 100000) / 100000;
-};
-
-const toNumber = (value: unknown, fallback = 0): number => {
-  if (typeof value === 'number' && Number.isFinite(value)) {
-    return value;
-  }
-  if (typeof value === 'string') {
-    const parsed = Number(value);
-    if (Number.isFinite(parsed)) {
-      return parsed;
-    }
-  }
-  return fallback;
-};
-
-const toStringArray = (value: unknown, fallback: string[]): string[] => {
-  if (!Array.isArray(value)) {
-    return fallback;
-  }
-  const filtered = value.filter(
-    (item): item is string => typeof item === 'string'
-  );
-  return filtered.length > 0 ? filtered : fallback;
-};
-
-interface EditModelFormProps {
-  model: Model;
-  providerId?: number;
-  onModelUpdate?: () => void;
-  onCancel?: () => void;
-  isOpen: boolean;
-}
-
-interface AdminModelData {
-  id: string;
-  name: string;
-  description?: string;
-  created: number;
-  context_length: number;
-  architecture: {
-    modality: string;
-    input_modalities: string[];
-    output_modalities: string[];
-    tokenizer: string;
-    instruct_type: string | null;
-  };
-  pricing: {
-    prompt: number;
-    completion: number;
-    request: number;
-    image: number;
-    web_search: number;
-    internal_reasoning: number;
-  };
-  per_request_limits: null | undefined;
-  top_provider: null | undefined;
-  upstream_provider_id: number;
-  enabled: boolean;
-}
-
-const normalizeAdminModelData = (
-  adminModel: AdminModel,
-  fallbackModel: Model,
-  providerId: number
-): AdminModelData => {
-  const pricingRecord =
-    adminModel.pricing && typeof adminModel.pricing === 'object'
-      ? (adminModel.pricing as Record<string, unknown>)
-      : {};
-
-  const architectureRecord =
-    adminModel.architecture && typeof adminModel.architecture === 'object'
-      ? (adminModel.architecture as Record<string, unknown>)
-      : {};
-
-  return {
-    id: adminModel.id,
-    name: adminModel.name,
-    description: adminModel.description || '',
-    created: toNumber(adminModel.created, Math.floor(Date.now() / 1000)),
-    context_length: Math.max(
-      0,
-      Math.trunc(
-        toNumber(adminModel.context_length, fallbackModel.contextLength || 4096)
-      )
-    ),
-    architecture: {
-      modality:
-        typeof architectureRecord.modality === 'string'
-          ? architectureRecord.modality
-          : fallbackModel.modelType || 'text',
-      input_modalities: toStringArray(architectureRecord.input_modalities, [
-        fallbackModel.modelType || 'text',
-      ]),
-      output_modalities: toStringArray(architectureRecord.output_modalities, [
-        fallbackModel.modelType || 'text',
-      ]),
-      tokenizer:
-        typeof architectureRecord.tokenizer === 'string'
-          ? architectureRecord.tokenizer
-          : '',
-      instruct_type:
-        typeof architectureRecord.instruct_type === 'string'
-          ? architectureRecord.instruct_type
-          : null,
-    },
-    pricing: {
-      prompt: roundToFiveDecimals(
-        toNumber(pricingRecord.prompt, fallbackModel.input_cost)
-      ),
-      completion: roundToFiveDecimals(
-        toNumber(pricingRecord.completion, fallbackModel.output_cost)
-      ),
-      request: toNumber(pricingRecord.request, 0),
-      image: toNumber(pricingRecord.image, 0),
-      web_search: toNumber(pricingRecord.web_search, 0),
-      internal_reasoning: toNumber(pricingRecord.internal_reasoning, 0),
-    },
-    per_request_limits:
-      adminModel.per_request_limits === null ||
-      adminModel.per_request_limits === undefined
-        ? adminModel.per_request_limits
-        : null,
-    top_provider:
-      adminModel.top_provider === null || adminModel.top_provider === undefined
-        ? adminModel.top_provider
-        : null,
-    upstream_provider_id:
-      typeof adminModel.upstream_provider_id === 'number'
-        ? adminModel.upstream_provider_id
-        : providerId,
-    enabled: adminModel.enabled !== false,
-  };
-};
-
-export function EditModelForm({
-  model,
-  providerId,
-  onModelUpdate,
-  onCancel,
-  isOpen,
-}: EditModelFormProps) {
-  const [isSubmitting, setIsSubmitting] = useState(false);
-  const [adminModelData, setAdminModelData] = useState<AdminModelData | null>(
-    null
-  );
-  const [isNewOverride, setIsNewOverride] = useState(false);
-
-  const form = useForm<EditModelFormData>({
-    resolver: zodResolver(EditModelFormSchema),
-    defaultValues: {
-      name: model.name,
-      description: model.description || '',
-      context_length: model.contextLength || 4096,
-      prompt: roundToFiveDecimals(model.input_cost),
-      completion: roundToFiveDecimals(model.output_cost),
-      enabled: model.isEnabled !== false,
-    },
-  });
-
-  const loadAdminModel = useCallback(async () => {
-    if (!providerId) {
-      console.error('loadAdminModel called without providerId');
-      return;
-    }
-
-    try {
-      const adminModel = await AdminService.getProviderModel(
-        providerId,
-        model.id
-      );
-
-      const normalizedAdminModel = normalizeAdminModelData(
-        adminModel,
-        model,
-        providerId
-      );
-      setAdminModelData(normalizedAdminModel);
-      setIsNewOverride(false);
-
-      form.reset({
-        name: normalizedAdminModel.name,
-        description: normalizedAdminModel.description || '',
-        context_length: normalizedAdminModel.context_length,
-        prompt: normalizedAdminModel.pricing.prompt,
-        completion: normalizedAdminModel.pricing.completion,
-        enabled: normalizedAdminModel.enabled !== false,
-      });
-    } catch {
-      setIsNewOverride(true);
-      setAdminModelData({
-        id: model.full_name,
-        name: model.name,
-        description: model.description || '',
-        created: Math.floor(Date.now() / 1000),
-        context_length: model.contextLength || 4096,
-        architecture: {
-          modality: model.modelType || 'text',
-          input_modalities: [model.modelType || 'text'],
-          output_modalities: [model.modelType || 'text'],
-          tokenizer: '',
-          instruct_type: null,
-        },
-        pricing: {
-          prompt: roundToFiveDecimals(model.input_cost),
-          completion: roundToFiveDecimals(model.output_cost),
-          request: 0,
-          image: 0,
-          web_search: 0,
-          internal_reasoning: 0,
-        },
-        per_request_limits: null,
-        top_provider: null,
-        upstream_provider_id: providerId,
-        enabled: model.isEnabled !== false,
-      });
-
-      form.reset({
-        name: model.name,
-        description: model.description || '',
-        context_length: model.contextLength || 4096,
-        prompt: roundToFiveDecimals(model.input_cost),
-        completion: roundToFiveDecimals(model.output_cost),
-        enabled: model.isEnabled !== false,
-      });
-    }
-  }, [providerId, model, form]);
-
-  useEffect(() => {
-    if (isOpen && providerId) {
-      loadAdminModel();
-    } else if (isOpen && !providerId) {
-      console.error('EditModelForm opened without providerId', {
-        model,
-        providerId,
-      });
-      toast.error('Missing provider information for this model');
-    }
-  }, [isOpen, providerId, model, loadAdminModel]);
-
-  const onSubmit = async (data: EditModelFormData) => {
-    if (!providerId) {
-      console.error('onSubmit called without providerId', {
-        model,
-        providerId,
-      });
-      toast.error('Missing provider ID - cannot update model');
-      return;
-    }
-
-    if (!adminModelData) {
-      console.error('onSubmit called without adminModelData', {
-        model,
-        providerId,
-        adminModelData,
-      });
-      toast.error('Model data not loaded - please try reopening the form');
-      return;
-    }
-
-    setIsSubmitting(true);
-    try {
-      const payload = {
-        id: adminModelData.id,
-        name: data.name,
-        description: data.description || '',
-        created: adminModelData.created || Math.floor(Date.now() / 1000),
-        context_length: data.context_length,
-        architecture: adminModelData.architecture || {
-          modality: 'text',
-          input_modalities: ['text'],
-          output_modalities: ['text'],
-          tokenizer: '',
-          instruct_type: null,
-        },
-        pricing: {
-          prompt: roundToFiveDecimals(data.prompt),
-          completion: roundToFiveDecimals(data.completion),
-          request: 0,
-          image: 0,
-          web_search: 0,
-          internal_reasoning: 0,
-        },
-        per_request_limits: adminModelData.per_request_limits,
-        top_provider: adminModelData.top_provider,
-        upstream_provider_id: providerId,
-        enabled: data.enabled,
-      };
-
-      if (isNewOverride) {
-        await AdminService.createProviderModel(providerId, payload);
-        toast.success('Model override created successfully!');
-      } else {
-        await AdminService.updateProviderModel(
-          providerId,
-          adminModelData.id,
-          payload
-        );
-        toast.success('Model updated successfully!');
-      }
-
-      onModelUpdate?.();
-      onCancel?.();
-    } catch (error) {
-      const action = isNewOverride ? 'create' : 'update';
-      toast.error(`Failed to ${action} model. Please try again.`);
-      console.error(`Error ${action}ing model:`, error);
-    } finally {
-      setIsSubmitting(false);
-    }
-  };
-
-  const handleClose = () => {
-    if (!isSubmitting) {
-      onCancel?.();
-    }
-  };
-
-  return (
-    <Dialog open={isOpen} onOpenChange={handleClose}>
-      <DialogContent className='max-h-[90vh] overflow-y-auto sm:max-w-[600px]'>
-        <DialogHeader>
-          <DialogTitle className='flex items-center gap-2'>
-            <Edit3 className='h-5 w-5' />
-            {isNewOverride ? 'Create Model Override' : 'Edit Model Override'}
-          </DialogTitle>
-          <DialogDescription>
-            {isNewOverride
-              ? `Create an override for &quot;${model.name}&quot;`
-              : `Update the model override for &quot;${model.name}&quot;`}
-          </DialogDescription>
-        </DialogHeader>
-
-        <Form {...form}>
-          <form onSubmit={form.handleSubmit(onSubmit)} className='space-y-4'>
-            <div className='grid grid-cols-1 gap-4 sm:grid-cols-2'>
-              <FormField
-                control={form.control}
-                name='name'
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Display Name *</FormLabel>
-                    <FormControl>
-                      <Input
-                        placeholder='e.g., GPT-4'
-                        {...field}
-                        className='w-full'
-                      />
-                    </FormControl>
-                    <FormDescription>
-                      Custom display name for the model
-                    </FormDescription>
-                    <FormMessage />
-                  </FormItem>
-                )}
-              />
-
-              <FormField
-                control={form.control}
-                name='context_length'
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Context Length *</FormLabel>
-                    <FormControl>
-                      <Input
-                        type='number'
-                        min='0'
-                        placeholder='4096'
-                        value={field.value ?? ''}
-                        onChange={(e) => {
-                          const value = e.target.value;
-                          field.onChange(
-                            value === '' ? 0 : parseInt(value, 10) || 0
-                          );
-                        }}
-                        onBlur={(e) => {
-                          const value = parseInt(e.target.value, 10);
-                          field.onChange(
-                            Number.isNaN(value) ? 0 : Math.max(0, value)
-                          );
-                        }}
-                        className='w-full'
-                      />
-                    </FormControl>
-                    <FormDescription>
-                      Maximum context window size
-                    </FormDescription>
-                    <FormMessage />
-                  </FormItem>
-                )}
-              />
-            </div>
-
-            <FormField
-              control={form.control}
-              name='description'
-              render={({ field }) => (
-                <FormItem>
-                  <FormLabel>Description</FormLabel>
-                  <FormControl>
-                    <Textarea
-                      placeholder='Brief description of the model...'
-                      {...field}
-                      rows={3}
-                      className='w-full'
-                    />
-                  </FormControl>
-                  <FormDescription>
-                    Optional description or notes about the model
-                  </FormDescription>
-                  <FormMessage />
-                </FormItem>
-              )}
-            />
-
-            <div className='grid grid-cols-1 gap-4 sm:grid-cols-2'>
-              <FormField
-                control={form.control}
-                name='prompt'
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Input Cost (per 1M tokens) *</FormLabel>
-                    <FormControl>
-                      <Input
-                        type='number'
-                        step='0.00001'
-                        min='0'
-                        placeholder='5.00000'
-                        value={field.value ?? ''}
-                        onChange={(e) => {
-                          const value = e.target.value;
-                          field.onChange(value === '' ? 0 : parseFloat(value));
-                        }}
-                        onBlur={(e) => {
-                          const value = parseFloat(e.target.value);
-                          field.onChange(roundToFiveDecimals(value));
-                        }}
-                        className='w-full'
-                      />
-                    </FormControl>
-                    <FormDescription>
-                      Cost in USD per 1,000,000 input tokens (max 5 decimals)
-                    </FormDescription>
-                    <FormMessage />
-                  </FormItem>
-                )}
-              />
-
-              <FormField
-                control={form.control}
-                name='completion'
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Output Cost (per 1M tokens) *</FormLabel>
-                    <FormControl>
-                      <Input
-                        type='number'
-                        step='0.00001'
-                        min='0'
-                        placeholder='15.00000'
-                        value={field.value ?? ''}
-                        onChange={(e) => {
-                          const value = e.target.value;
-                          field.onChange(value === '' ? 0 : parseFloat(value));
-                        }}
-                        onBlur={(e) => {
-                          const value = parseFloat(e.target.value);
-                          field.onChange(roundToFiveDecimals(value));
-                        }}
-                        className='w-full'
-                      />
-                    </FormControl>
-                    <FormDescription>
-                      Cost in USD per 1,000,000 output tokens (max 5 decimals)
-                    </FormDescription>
-                    <FormMessage />
-                  </FormItem>
-                )}
-              />
-            </div>
-
-            <FormField
-              control={form.control}
-              name='enabled'
-              render={({ field }) => (
-                <FormItem className='flex flex-row items-center justify-between rounded-lg border p-4'>
-                  <div className='space-y-0.5'>
-                    <FormLabel className='text-base'>Model Enabled</FormLabel>
-                    <FormDescription>
-                      Enable or disable this model override
-                    </FormDescription>
-                  </div>
-                  <FormControl>
-                    <Switch
-                      checked={field.value}
-                      onCheckedChange={field.onChange}
-                    />
-                  </FormControl>
-                </FormItem>
-              )}
-            />
-
-            <div className='flex justify-end gap-2 pt-4'>
-              <Button
-                type='button'
-                variant='outline'
-                onClick={handleClose}
-                disabled={isSubmitting}
-              >
-                Cancel
-              </Button>
-              <Button type='submit' disabled={isSubmitting}>
-                {isSubmitting ? (
-                  <>
-                    <Loader2 className='mr-2 h-4 w-4 animate-spin' />
-                    {isNewOverride ? 'Creating...' : 'Updating...'}
-                  </>
-                ) : isNewOverride ? (
-                  'Create Override'
-                ) : (
-                  'Update Model'
-                )}
-              </Button>
-            </div>
-          </form>
-        </Form>
-      </DialogContent>
-    </Dialog>
-  );
-}

ui/components/providers/RoutstrCreateKeySection.tsx

@@ -1,6 +1,6 @@
 'use client';
 
-import { useCallback, useState } from 'react';
+import { useState } from 'react';
 import Image from 'next/image';
 import { Copy, Loader2, Zap, KeyRound } from 'lucide-react';
 import { toast } from 'sonner';
@@ -10,7 +10,6 @@ import { Input } from '@/components/ui/input';
 import { Textarea } from '@/components/ui/textarea';
 import { Label } from '@/components/ui/label';
 import { Badge } from '@/components/ui/badge';
-import { Separator } from '@/components/ui/separator';
 import { Tabs, TabsContent, TabsList, TabsTrigger } from '@/components/ui/tabs';
 
 interface RoutstrCreateKeySectionProps {

ui/lib/api/services/admin.ts

@@ -76,6 +76,7 @@ export const AdminModelSchema = z.object({
   canonical_slug: z.string().nullable().optional(),
   alias_ids: z.array(z.string()).nullable().optional(),
   enabled: z.boolean().default(true),
+  forwarded_model_id: z.string().nullable().optional(),
 });
 
 export const ProviderModelsSchema = z.object({