A Model List Does Not Prove a Codex Relay Supports the Responses API

#ai #openai #api #testing

An OpenAI-compatible /v1/models response is useful discovery evidence, but it
does not prove that the same key can complete a Codex workflow. Model discovery,
Chat Completions, the Responses API, streaming, tool calls, and client-specific
gateways may use different routes or permissions.

AI Relay Check therefore records compatibility as a specific combination of:

provider;
model ID;
recorded protocol;
test key and route;
timestamp;
generation outcome.

Current evidence scope

The current public Codex guide contains GPT-family samples from six paid test
accounts. Three providers have samples recorded with the Responses protocol.
Across those Responses-labeled records, 50 of 64 requests completed
successfully. The latest included GPT sample is dated October 1, 2026.

These figures are dated test samples. They are not uptime percentages and do not
prove that every Codex client feature or route group works.

Five checks before choosing a Codex relay

1. Confirm the exact base URL

Use the HTTPS base URL assigned to the tested product and route group. A
provider may expose different endpoints for a general OpenAI-compatible API,
Codex-only keys, subscriptions, and pooled routes.

2. Test the generation endpoint

Determine whether the route uses /v1/responses,
/v1/chat/completions, or a client-specific gateway. A model-list request can
pass while the required generation endpoint returns 400, 404, 429, or 503.

3. Fix the model ID and key group

Run the exact model needed by the project and record the group assigned to the
test key. A successful request from one group should not be attributed to a
different provider-published price group.

4. Verify streaming and tools separately

A basic text response does not establish streaming, tool calls, structured
output, or the complete Codex client workflow. Test the features the application
actually depends on.

5. Keep compatibility and billing separate

Preserve returned token fields and compare a controlled request with the
provider's account meter or wallet. A successful response does not prove that
the published input, output, cache, or subscription rate was applied.

Public evidence

Conclusion

The right question is not whether a relay lists a GPT model. The useful question
is whether the exact model, key group, generation endpoint, and required Codex
features completed under recorded conditions. Billing and model identity still
require separate evidence.

Disclosure: AI Relay Check has referral relationships with some tested
providers. Referral status is excluded from request outcomes and ranking
weights. Unverified model identity, route provenance, and billing remain
explicitly labeled.

DEV Community