Gate 1
Scope and rate card
Pin the model, default service tier, official prices, verification date, input and output limits, cache categories, run owner, maximum budget, and excluded charges.
Reserve, settle, reconcile
OpenAI Agent Run Budget Readiness
Describe one Responses API run, its verified rate card, request limits, tools, concurrency, and failure handling. Generate a checklist without mistaking an application estimate for the final provider bill.
Four readiness gates
Gate 2
Reservation and concurrency
Count the complete request, reserve worst-case token cost atomically, use unique holds, prevent double reservation across workers, and reject calls that exceed the balance.
Gate 3
Settlement and blocking
Validate model, tier, status, token details, totals, bounds, and actual cost; release unused funds only after settlement and block permanently on timeout or uncertain charge.
Gate 4
External costs and reconciliation
Price or prohibit hosted tools, streaming, background work, non-default tiers, regional and long-context rates, storage, retries, and third parties; reconcile provider usage and invoices.
Classify the control
Ready for offline tests
The state machine, exact arithmetic, holds, failures, and fixtures are implemented with fictional data and no network calls.
Ready for a bounded API pilot
The live rate card is verified, unsupported costs are prohibited, retries are controlled, concurrency is atomic, and every uncertainty blocks the run.
Not a hard spend cap
The ledger is still an application estimate; project controls, provider usage, late charges, invoice reconciliation, and incident ownership remain required.
Minimum acceptance evidence
Prove an exact-budget success, a blocked over-budget next call, atomic competing reservations, timeout blocking, missing-usage blocking, incomplete-response settlement, disabled or separately reserved retries, prohibited unsupported charges, and reconciliation against provider usage. Run failure fixtures offline before any paid call.
Continue with the implementation guide
Responses API per-run spending controllerReview exact rate cards, input counting, worst-case holds, usage settlement, uncertain charges, external costs, and provider reconciliation.Official pattern checked August 19, 2026 against the OpenAI Cookbook. Recheck SDK, model, tier, prices, token counting, usage fields, tools, retries, limits, and billing before deployment.
Frequently Asked Questions
It creates a checklist from the run budget, rate card, token bounds, reservation, settlement, concurrency, failure, external-charge, and reconciliation facts you provide. It does not call the API or enforce a provider billing limit.
No. OpenAI states that its sample model, prices, token counts, and limits are fictional. Verify the current official model and pricing pages and record when that check occurred.
The actual input mix and output length are not known before the call. Reserving the highest plausible input category plus maximum output prevents two requests from spending the same remaining budget.
Block the run and keep the reservation. A timeout does not prove processing stopped, so reconcile provider usage before authorizing any further call.
No. Alerts notify, and project spending limits may not take effect immediately. Keep the run controller and provider reconciliation as separate controls.