Model selection, documented
Every request
gets a call sheet.
Catwalk is an OpenAI-compatible gateway that filters the field against your constraints, records the selection, and attaches the decision to the wallet debit.
Inspect the casting decision ↓- OpenAI-compatible request shape
- Hard budget constraints
- Integer micro-USD accounting
- Receipt per completed run
The full decision field
Illustrative selection / run_demo_01
Brief
Business writing best_value max $0.05 quality floor 0.85- openai/gpt-5.4-mini Quality score 0.84 is below the displayed 0.85 floor Cut
- google/gemini-3.1-pro-preview Eligible candidate Hold
- mistral/mistral-medium-2508 Highest illustrative score after filters Cast
- anthropic/claude-sonnet-5 Eligible candidate Hold
The operating principle
The selection only matters if you can audit it.
A completed run connects the objective and hard filters to the selected model, recorded usage, latency, fallback attempts, and wallet debit.
Scorecards are synthetic today. Catwalk does not present “best value” as measured provider quality until the live evaluation harness has run.
How the assignment moves
One request in.
One documented decision out.
-
Brief
Keep the request shape
Use an OpenAI-compatible client with a Catwalk base URL and key.
client = OpenAI( base_url="https://your-catwalk-host.example/v1", api_key="catwalk_…" ) -
Field
Set the casting brief
Pin a model or state the objective and non-negotiable limits.
"catwalk": { "objective": "best_value", "max_cost_usd": 0.05, "quality_floor": 0.85 } -
Call sheet
Receive work plus evidence
The response remains OpenAI-shaped and links to its routing receipt.
"catwalk": { "routing_reason": "…", "actual_cost_usd": 0.0024, "receipt_id": "rcpt_…" }
Receipt attached
Explainability that reaches the ledger.
Each completed run connects the routing decision to estimated cost, provider-reported usage, wallet debit, latency, and fallback attempts.
- Why the selected model remained eligible
- Which constraint cut each excluded candidate
- Estimated cost compared with recorded actual cost
- Every bounded fallback attempt
- run
- run_demo_01
- objective
- best_value
- selected / Cast
- mistral/mistral-medium-2508
- reason
- best value above the displayed quality floor
- Cut
- 6 chat candidates / quality_floor
- Hold
- 4 eligible alternatives
- fallbacks
- none
- estimated cost
- $0.0024*
- recorded actual
- $0.0024*
- wallet debit
- $0.0024*
The commercial line item
5.5% service fee
The service fee is added to the wallet credit purchased; it does not reduce the credit added to your wallet. The fee is 5.5% with a $0.80 minimum. Usage then debits the wallet at the price stored in Catwalk's model catalog.
- Wallet credit
- $100.00 wallet credit
- Service fee
- + $5.50
- Total charged
- $105.50
Readiness ledger
Built in code.
Not yet proven in production.
Implemented + tested
- Constraint filtering and objective-based routing
- Prepaid wallet with a never-negative hard cap
- Receipts, fallback records, and spend controls
- Chat, image, video, speech, and transcription paths
Required before launch
- Live smoke tests across the current provider catalog
- Managed Postgres and Redis in staging
- Live Stripe webhook and end-to-end paid transaction
- Security, observability, backup, and legal review
Core routing, wallet, receipt, and modality paths are implemented and covered by deterministic tests. Catwalk is not presented as production-ready.
Buyer review
Questions worth putting on the call sheet.
Will every OpenAI integration work unchanged?
Catwalk implements OpenAI-compatible request and response shapes for its supported endpoints, including chat streaming. Provider-specific SDK features outside that surface may require changes.
How does “best value” know what quality means?
The router uses catalog scorecards plus the task, objective, and constraints. Current scorecards are synthetic and must not be treated as measured provider quality.
Are provider prices passed through exactly?
Usage debits use prices stored in Catwalk's catalog. Current model IDs, rates, and usage accounting still require live smoke tests. The service fee is separately added to the wallet-credit purchase at 5.5%, subject to a $0.80 minimum.
Does Catwalk store prompt content?
Structured logs redact prompt content by default. Productized opt-in and retention controls described in the specification are not yet built.
Private beta
Review the mechanism.
Keep the caveats attached.
Evaluation is prepaid. A 5.5% service fee, subject to a $0.80 minimum, is added to each wallet-credit purchase. Inspect the implementation before funding a wallet.