GPT-5.5 Leaves Codex on October 14 — but Not the API

The vendor will retire GPT-5.5 from ChatGPT, ChatGPT Work and Codex on October 14, 2026 — and the single most important word in that sentence is those. The model does not disappear. It stays on the vendor's API platform, and it stays inside Codex sessions that authenticate with your own API key rather than a ChatGPT sign-in. Two customers on the same product, in the same week, will get opposite answers about whether their work still runs.

That split is the whole story, and most of the coverage published today has flattened it into "OpenAI is discontinuing GPT-5.5." It is worth being precise about what is actually ending, because the difference decides whether you spend the next four weeks migrating or do nothing at all.

What actually changes on October 14

• ChatGPT, ChatGPT Work and Codex on a ChatGPT sign-in — GPT-5.5 is removed on all plans, consumer through Enterprise and Edu.

• The OpenAI API Platform — unaffected. GPT-5.5 is not listed on OpenAI's own deprecations page for any API shutdown date.

• Codex authenticated with your own API key — unaffected, for the same reason: the retirement is scoped to the subscription surfaces, not to the model.

• Your workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts — these are the things that actually break, because they select a model by ID and will keep asking for one that no longer answers on those surfaces.

OpenAI's guidance for anyone in the affected group is to move to GPT-5.6 Sol or GPT-6 Astra before the deadline. Both are live, and Codex users on a ChatGPT sign-in are being pointed at gpt-5.6-sol specifically. If you are in that group, the work is not "evaluate a new model" — it is "find every place a model ID is hardcoded and change it." That is a grep problem, and it is worth doing this week rather than on October 13.

Why the aggregators get this wrong

Within hours of the announcement, a wave of secondary coverage compressed the retirement into a single date and a single scope. At least one widely-copied aggregation placed gpt-5.5 in OpenAI's separate legacy-API deprecation wave with an October 23 shutdown date. That wave is real — it is the one taking gpt-4, o1, o1-pro, o3-mini, gpt-4o-2024-05-13 and a long tail of fine-tuned variants off the API — but GPT-5.5 is not in it. Check OpenAI's deprecations page and you will not find gpt-5.5 or gpt-5.5-pro anywhere on it.

The confusion is understandable. OpenAI is running several retirements at once in 2026, and they have different dates, different scopes and different replacement models. The practical lesson for anyone building on this platform: read the surface, not the headline. "Discontinued" in a news summary can mean the chat product, the API, or a specific authentication path, and those are three different migrations.

The Codex distinction is the one that catches people

Codex is where this gets genuinely confusing, because Codex is one product with two ways in. Sign in with your ChatGPT account and you are spending plan capacity against a subscription surface — that path loses GPT-5.5 on October 14. Sign in with an API key and you are spending API credit against the API platform — that path keeps it.

This is not a new pattern. When GPT-5.2 and GPT-5.2-Codex were sunset earlier in 2026, ChatGPT-authenticated CLI users lost access to the model in the TUI selector on June 2, while API-key users kept working against the same model until the later API-side shutdown dates in July and August. The subscription surface has always moved first. If you have ever run codex login --with-api-key, you are on the path that survives this one.

One caveat worth stating plainly, because it has bitten routing setups before: OpenAI tightened the Codex responses-lite endpoint in late June, and gateways that forward an internal lite header have been getting 400 unsupported_value back for GPT-5.5 ever since. If you reach GPT-5.5 through a proxy rather than directly, test that path before you assume the October 14 change is the only thing that could interrupt you.

What you are keeping, in numbers

For anyone weighing whether the surviving API path is worth keeping in a stack, GPT-5.5's current terms are unchanged by the announcement. It is priced at $5.00 per million input tokens and $30.00 per million output tokens, with cached input reads at $0.50 and a long-context tier that changes rates above 272,000 input tokens. The Pro tier runs $30.00 and $180.00. Context is roughly 922K input and 128K output, accepting text, image and file input.

The independent picture is already cooling, though, and that is worth knowing before you plan around it. Artificial Analysis now flags GPT-5.5 as deprecated on its own leaderboard and has wound its testing back to default 10,000-token workloads — the same signal a product retirement sends, arriving from the evaluation side. Its current Intelligence Index reading for the model is 39 on the v4.3 index, with throughput around 94 output tokens per second and a time-to-first-token of about 35 seconds, which sits well above the median for comparable reasoning models. OpenAI's own reported figures — Terminal-Bench 2.0 at 82.7%, SWE-bench Verified at 82.6%, SWE-bench Pro at 58.6% — remain vendor-reported and unreproduced by a neutral lab.

The migration is smaller than it looks from our side

If you route through one API rather than contracting directly with each vendor, this is a config change and not a project. GPT-5.5 and its successor GPT-5.6 Sol are both in our catalogue — as is GPT-5.5 Pro — so moving a workload from one to the other is a string in a request body against a single key, with no second contract and no integration work. Our catalogue currently lists 196 models across 15 providers, priced at provider list rate with 0% markup, which means the numbers above are the numbers you pay and any vendor price change lands the same day.

That structure also happens to be the right shape for a retirement. When a model ID stops answering on a surface you do not control, automatic failover is what keeps the request alive, and the routing DSL is what lets you split traffic — a percentage to the successor, the remainder held on the incumbent — so you can measure the migration on your own workload instead of trusting a vendor's framing of it.

What to watch

Two dates matter and neither is October 14. The first is whenever OpenAI adds GPT-5.5 to the actual API deprecations page — that, not today's announcement, is the one that ends the API path, and it will come with a notice period rather than a surprise. The second is whether Artificial Analysis restores full benchmarking or retires the entry entirely, because once independent evaluation stops, every remaining GPT-5.5 number in circulation is a vendor number.

The honest summary for this week: if you are on Codex with a ChatGPT sign-in, you have a real migration and it is due October 14. If you call GPT-5.5 through the API or an API-key Codex session, nothing changed today, you are not on the deprecation list, and the correct response is to note the date and keep building.

The surviving path is a single endpoint — GPT-5.5 stays callable there with the same key that reaches GPT-5.6 Sol.