ChatGPT subscription

The chatgpt provider calls the Codex backend that ships with a ChatGPT subscription (Plus, Pro, Business, or Enterprise). Usage is billed against the subscription’s quota, not an OpenAI Platform API key. Pair it with the Codex guide to run Codex -> GoModel -> ChatGPT subscription.

Configure

The credential is the access token from your Codex sign-in:

CHATGPT_API_KEY=$(jq -r .tokens.access_token ~/.codex/auth.json)

Or in config.yaml:

providers:
  chatgpt:
    type: chatgpt
    api_key: "${CHATGPT_API_KEY}"

Run codex login first if ~/.codex/auth.json does not exist yet. GoModel derives the ChatGPT account ID from the token itself, so nothing else is needed.

The token expires roughly every 10 days. Re-export it (the Codex CLI refreshes the file whenever it runs) or run codex login again when requests start returning 401.

Models

The Codex backend has no model-listing endpoint, so GoModel ships the inventory a ChatGPT subscription can call:

gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5

Override it when your plan serves a different set:

CHATGPT_MODELS=gpt-5.6-sol,gpt-5.6-terra

A model outside the plan’s set is rejected upstream with The '<model>' model is not supported when using Codex with a ChatGPT account. gpt-5.4 and gpt-5.4-mini leave ChatGPT-authenticated Codex on August 31, 2026; use gpt-5.6-terra and gpt-5.6-luna instead. Both stay available to Codex sessions authenticated with an OpenAI API key, through the openai provider.

Supported surfaces

The Codex backend serves /responses and nothing else. GoModel translates /v1/chat/completions onto it — request, response, and streaming — so chat clients work too. /v1/embeddings still answers 501: the backend has no embeddings endpoint. Chat requests are converted to Responses requests, executed upstream, and converted back to chat completions. The returned chatcmpl- IDs are labels, not resource handles: GoModel pins store: false and drops previous_response_id (the backend allows neither), so no chaining is possible and chat clients resend full history as usual. Chat parameters with no Responses equivalent are rejected with a 400 that names the field, before any upstream call: n other than 1, logit_bias, stop, seed, frequency_penalty, presence_penalty, logprobs, top_logprobs, modalities, audio, web_search_options, and the deprecated functions / function_call pair. Zero spellings of logprobs, top_logprobs, and the penalties (false / 0) are tolerated: they change nothing. prediction is dropped instead: it is a speed hint and never changes the answer. The backend also validates against a strict parameter allowlist. GoModel adapts requests rather than failing them, so callers keep using the standard Responses API:

Request field	Behavior
`stream`, `store`	Pinned to `true` / `false` — the backend rejects anything else
`input` as a string	Wrapped into the message list the backend requires
`instructions`, `tools`, `tool_choice`, `parallel_tool_calls`, `reasoning`, `text`, `include`	Forwarded
`temperature`, `top_p`, `max_output_tokens`, `previous_response_id`, `truncation`, `metadata`, `user`, `service_tier`, `top_logprobs`	Dropped — unsupported upstream

Translated chat requests go through the same allowlist: temperature, top_p, and max_tokens / max_completion_tokens map to valid Responses fields and are then silently dropped upstream, exactly as for native Responses requests. Because the backend streams only, a non-streaming POST /v1/responses is served by streaming upstream and returning the final response object. Clients see a normal non-streaming response. See Responses compatibility for how the gateway’s Responses surface behaves across providers.

Reported cost is not real spend

Subscription usage is flat-rate, but these model IDs also exist on the OpenAI Platform, so the model catalog attaches their per-token API prices. Usage records and dashboard totals for chatgpt show a figure that corresponds to no actual charge, and the same collision makes GET /v1/models advertise modes: ["chat", "responses"]. The modes are cosmetic — they drive dashboard grouping, not routing. Pricing is not: it feeds cost tracking, budgets, and cost load balancing.

A budget can therefore reject chatgpt traffic for “spending” money the subscription never charges. Scope budgets to a user path that excludes subscription traffic, or leave budgets off for it. Declaring per-model pricing and modes: ["responses"] in model metadata corrects what /v1/models advertises, but usage records still price at catalog rates. Declare every model you want served: the list replaces the default inventory rather than adding to it.

Limits

Subscription quota is separate from API credit. When it is exhausted the gateway relays a 429:

{"error":{"message":"The usage limit has been reached","type":"rate_limit_error"}}

OpenAI authorizes these tokens for Codex traffic. Use this provider to route your own Codex usage through GoModel, not as a general-purpose inference API.

​Configure

​Models

​Supported surfaces

​Reported cost is not real spend

​Limits

Configure

Models

Supported surfaces

Reported cost is not real spend

Limits