Troubleshooting

Start with evidence

Record the HTTP status, error code, time, region, platform model ID, and returned request/response identifiers. Check . Remove full keys and sensitive request content before sharing logs.

Symptom	Check
401 / invalid key	Full key, disabled/revoked/expired state, protocol-specific headers
403 / denied access	Model permissions on the key and account status
Insufficient balance or budget	Available balance, key spending limit, output budget, active reservations
404 / model or path missing	Platform model ID, regional URL, repeated `/v1`, required protocol
429 / rate limited	Gateway/upstream limits, concurrency, retry hints; confirm execution first
502 / 503 / 504	Route health, channel capabilities, upstream service, timeouts, invalid usage
Coding client cannot connect	Active provider, protocol, model capabilities, environment variables
Empty model catalog	Administrator model/channel/route configuration and enablement

Interruptions and cancellation

A disconnect does not prove that nothing ran. Query a stored Responses request through the original region; otherwise check request records. Execution state and actual usage may arrive separately. Cancellation does not establish zero charges.

Retries and support

Disable automatic SDK retries. Submit a new request only after confirming the previous one did not execute or explicitly accepting another generation. A manually selected region does not silently switch to another region. Give administrators sanitized identifiers, time, and region, never a complete API key. Platform records determine permissions and charges.

Getting started

SDKs

Coding tools

Resources

Start with evidence

Interruptions and cancellation

Retries and support

​Start with evidence

​Interruptions and cancellation

​Retries and support

Start with evidence

Interruptions and cancellation

Retries and support