On October 14, GPT-5.5 retires from Codex on every ChatGPT plan. If you sign into Codex with your ChatGPT account, your default model is gone in three days — and OpenAI's migration note points you at exactly one destination: another OpenAI model.
There's a second, documented door the announcement skips. Codex supports third-party model providers through config.toml, and any endpoint that speaks the Responses API qualifies — including open-weight models at roughly a quarter of the flagship's output price. Here's the official checklist, the option it skips, and what each one costs on a real bill.
Disclosure: Wallaby Token sells API access to Kimi K3, one of the third-party options below, and the endpoint in the setup section is ours. Everything attributed to OpenAI below is quoted from its Codex models documentation, fetched October 11, 2026.
What's actually retiring
The scope, per OpenAI's note: on October 14, 2026, GPT-5.5 leaves ChatGPT, ChatGPT Work, and Codex on all plans: consumer, Business, Enterprise, and Edu. Two boundaries matter more than the headline:
- The OpenAI API is not affected. GPT-5.5 stays available there.
- Codex authenticated with your own API key is not affected. Only ChatGPT sign-in loses the model.
So if you already run Codex on an API key, nothing in this post is urgent. The rest is for everyone who signed in with a ChatGPT account and let the defaults ride.
The official checklist
OpenAI's migration note lists six places where gpt-5.5 can hide: workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks, and scripts that select a model. Grep all six.
The desktop app, CLI, and IDE extension share one config.toml. To pin a replacement yourself:
model = "gpt-6.1-sol"
If you set nothing, Codex falls back to a recommended model.
Which replacement depends on your plan. On Plus, Pro, Business, Enterprise, and Edu, the note points to GPT-6 Sol; on Free and Go, to GPT-6 Luna in the desktop app. For complex coding and agentic work, OpenAI recommends GPT-6.1 Sol, which has been rolling out since September 29. Two caveats on that one: on Enterprise and Edu it stays off until an administrator enables it, and Free and Go are not included at launch. Availability depends on the rollout, your sign-in method, and your client, so check the model picker before assuming.
The door the announcement skips
Codex's CLI and IDE extension also load custom model providers from a [model_providers] block in config.toml. Any endpoint that speaks the OpenAI Responses API natively qualifies.
That qualifier is load-bearing. Since Codex 0.137.0, wire_api = "chat" is gone: current releases reject Chat Completions-only endpoints outright, and any tutorial written before the cutover will hard-error at startup. That rules out most generic OpenAI-compatible proxies — the backend must implement the full Responses API, and plain /v1/chat/completions compatibility is no longer enough for current Codex.
This route is documented, not hacky. Moonshot AI published its own Codex guide on September 28 (kimi.ai/resources/codex-api), walking through the same configuration for Kimi K3, down to model_context_window = 1048576 to match the model's 1M-token context. When the model vendor writes the guide, the second door is a supported entrance.
What it costs on a real bill
Per 1M tokens, as of October 11, 2026:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| Kimi K3 (official list) | $3.00 | $0.30 | $15.00 |
| Kimi K3 on Wallaby | $2.70 | $0.27 | $13.50 |
Wallaby's rate is 10% off the official list, ongoing.
Two things the table can't show. First, how often the cached column decides the bill: agent loops resend most of their context every turn, so cached-input pricing frequently dominates the real total. Second, output dominates the rate card: a coding agent reads far more than it writes, but each output token costs five times an input token on the K3 rows. For the full cross-provider picture on Kimi K3 (18 providers, blended at a coding-agent workload), see our pricing comparison.
When to stay official, when to go third-party
Stay official if your team runs on managed configurations, if Enterprise or Edu compliance is the reason Codex is allowed at all, or if you want one bill and zero extra vendors. The official replacements are good models, and the migration is one line of TOML.
Go third-party if you're an individual watching output-token costs, if you want an open-weight model you can also run elsewhere, or if you already manage API keys and don't want a subscription tied to a seat. The trade is real: billing, privacy, and troubleshooting move to the provider you pick, and a third-party provider won't get Codex-specific feature updates or official support from OpenAI — so pick one whose metering you can inspect. We itemize every request in the dashboard for exactly that reason, but the principle holds whoever you choose.
Setup in five minutes
The minimal working configuration — the same block Moonshot's own guide ships on current Codex, and the layout we verified end to end in our September walkthrough:
model = "kimi-k3"
model_provider = "wallaby"
[model_providers.wallaby]
name = "Wallaby"
base_url = "https://api.wallabytoken.com/v1"
env_key = "WALLABY_API_KEY"
wire_api = "responses"
Save it as ~/.codex/wallaby.config.toml, export WALLABY_API_KEY, and launch with codex --profile wallaby. Moonshot's guide puts the equivalent block directly in the main config.toml instead; both layouts work on current Codex. One detail worth getting right: the model value must be the model ID your provider actually serves — kimi-k3 here — not a display name; a mismatch fails with a model-not-found error.
The full walkthrough is in our earlier guide: Kimi K3 in Codex CLI: profile-file setup. It covers the harmless "model metadata not found" warning, the slow first turn, the apply_patch retry, and what to do when a stream disconnects, all with screenshots. Ten minutes there beats an hour of guessing here.
One more date
If GitHub Copilot is your driver instead of Codex, its GPT-5.5 retires on October 19 under a different replacement mapping — that one affects the Copilot-powered IDE experience, a different product from the standalone Codex CLI. Same discipline applies: read the official note, check every place a model is pinned, then decide.
A note on dates: this guide reflects Codex changes effective October 14, 2026, with sources fetched October 11. Model availability, pricing, and configuration requirements move fast — cross-check the official Codex documentation before acting on any of it.
If you want to test the third-party route before committing, a new Wallaby Token account comes with $0.50 in trial credit, enough to run Codex's full agent loop on a real task several times over. The minimum top-up is $20, and the rate stays 10% off the official list, ongoing.