When logged in with ChatGPT Plus/Pro while also supplying an API key, automatically fall back to using API credits once ChatGPT credits run out

What feature would you like to see?

When logged in with ChatGPT Plus/Pro while also supplying an API key, automatically fall back to using API credits once ChatGPT credits run out.

This along with greater visibility of used/remaining credits in one's ChatGPT plan would be killer.

Are you interested in implementing this feature?

Not unless I can vibe code it...

Additional information

Just for fun...

user
Would it be difficult to implement the auto-fallback from ChatGPT plan credits to API credits?

codex (gpt-5-high)
Short answer: Moderate effort, low risk. It’s localized and straightforward to add.

Why it’s feasible

• Single choke point: Requests go through codex-rs/core/src/client.rs::ModelClient::stream_responses, which already parses 429 errors and classifies usage_limit_reached / usage_not_included.
• Both creds available: auth.json can contain both ChatGPT tokens and OPENAI_API_KEY; env var OPENAI_API_KEY works too. We can read either without changing formats.
• URL selection handled: Switching auth automatically flips base URL (ChatGPT backend vs OpenAI API) via ModelProviderInfo::get_full_url.

Proposed approach

• Hook: In stream_responses, on 429 with error.type == "usage_limit_reached" or "usage_not_included" and current auth is ChatGPT:
• Try to resolve an API key:
• Env: std::env::var("OPENAI_API_KEY").
• File: codex_login::try_read_auth_json(get_auth_file(&self.config.codex_home))?.OPENAI_API_KEY.
• If found, switch a mutable local auth to Some(CodexAuth::from_api_key(key)), then retry the request once.
• Details to adjust:
• Recompute store per attempt: let store = prompt.store && auth.as_ref().map(|a| a.mode) != Some(AuthMode::ChatGPT).
• Only attempt fallback once (guard with a boolean).
• Keep current retry/backoff behavior for other errors.
• Continue sending chatgpt-account-id header only when auth.mode == ChatGPT (already conditional).

Edge cases

• No API key available: Return current error unchanged.
• Invalid/expired API key: Normal retry/UnexpectedStatus paths apply; no infinite loops.
• Custom OPENAI_BASE_URL: Respected; URL comes from provider config regardless of auth.
• Responses vs Chat wire API: Built‑in OpenAI uses Responses; fallback lives in stream_responses. If needed later, mirror logic in chat_completions path.

Tests to add

• Fallback success: First response returns 429 usage_limit_reached with ChatGPT headers; second succeeds with Bearer sk-test-key and no chatgpt-account-id.
• No key present: Still surfaces usage-limit error.
• One-shot: Ensure only one fallback attempt occurs (two total requests).
• Store flag: Assert store=false on first attempt (ChatGPT), store=true on fallback (API key).

Estimated changes

• Files: core/src/client.rs (+ ~40–80 LoC), core/tests/client.rs (+ 1–2 wiremock tests).
• Docs: Note auto-fallback behavior in README and optionally config.md (and add an opt-out flag later if desired).
• Effort: ~1/2 day including tests and docs.

If you’d like, I can implement this behind a default-on behavior (or add auto_fallback_to_apikey config) and ship tests.