GPT Codex Models: History, Prices and What Replaced Them

OpenAI's Codex-branded models are mostly gone. As of 31 October 2026, codex-mini-latest, GPT-5-Codex, GPT-5.1-Codex (and Max and Mini) and GPT-5.2-Codex are shut down, and GPT-5.3-Codex is deprecated with removal set for 1 April 2027. Codex now runs on the general GPT-6 models, led by Sol 6.1 at $2 in and $10 out per million tokens.

The Codex models at a glance

OpenAI trained a series of models for agentic coding in its Codex tools. Each page on developers.openai.com lists its status, context, price and cutoff.

Every row comes from the model's page on developers.openai.com, fetched 31 October 2026, and the deprecations page. All accept text and image input and return text, with reasoning-token support. GPT-5.2-Codex and GPT-5.3-Codex support low, medium, high and xhigh reasoning effort. Most of these models are available only through the Responses API, not Chat Completions.

The current Codex-line price points are a long way from today's. GPT-5.3-Codex at $1.75 and $14 compares with GPT-6.1 Sol at $2 and $10, so a move to the replacement raises input cost slightly and lowers output cost by about a quarter, before reasoning tokens.

Who makes them and the licence

OpenAI makes all of them. None has public weights or an open licence. They are API-only models, and several are now gone even from the API. If you want open weights from OpenAI, the only option is gpt-oss, which is a separate, Apache-2.0 line and not Codex-tuned.

Shutdown dates and replacements

OpenAI's deprecations page gives a clear migration map, and it matters if old model names linger in scripts.

The deprecations page sets a policy for this category. It says specialised variants of general models, with Codex variants as the named example, get at least three months' notice. GPT-5.3-Codex got six months. Note that the codex-mini-latest shutdown also ended the legacy local shell tool that only that model supported, per the same page.

Separately, GPT-5.5 retired from Codex with ChatGPT sign-in on 14 October 2026, per our Codex pricing page. API-key sessions were not affected. We found no OpenAI statement of a Codex-branded model for the GPT-6 line.

What Codex runs on now

OpenAI did not publish Codex-suffixed models for the GPT-5.6 or GPT-6 generations. Its models page describes GPT-6.1 Sol as being for "complex coding, computer use, and professional work" and GPT-6 Sol as "built to power complex coding and agentic workflows." The Codex pricing page lists these models with message allowances.

The figures come from OpenAI's Codex pricing page, as recorded on our Codex pricing page. OpenAI warns that allowances depend on task size and whether you run locally or in the cloud, so read them as ranges. Plus lists Sol 6.1 and Luna, and the Pro tiers add more. For API-key use, you pay token prices from the table on GPT models.

One more difference from the older line: Sol 6.1's model page says it uses a default context window of 500K tokens in Codex and ChatGPT Work, and subscription usage has no long-context surcharge there. The older Codex models had a 400,000-token window.

How OpenAI described each step

OpenAI's models page gives one line per model, and reading them in order shows how the branch developed.

Three patterns stand out. First, each Codex model was a tuned version of a general model released at the same time, so the Codex branch always trailed or matched the general line. Second, OpenAI offered a Mini variant at one tenth to one fifth of the flagship price, which is the same split the GPT-6 line now makes with Luna. Third, the context window rose from 200,000 on codex-mini-latest to 400,000 on the 5.x Codex models and to 1,050,000 on GPT-6, while max output rose from 100,000 to 128,000.

The knowledge cutoffs also moved. The 5.1 models stop at 30 September 2024, while 5.2-Codex and 5.3-Codex reach 31 August 2025 and the GPT-6 models reach April to June 2026. For a coding agent that must know recent library releases, that gap can matter more than a few benchmark points.

A cost comparison for one agent task

Our arithmetic from the listed rates, not a measurement. Take an agent task that reads 120,000 input tokens in total and writes 8,000 output tokens, before caching and reasoning tokens.

On those numbers, moving from the old Codex model to Sol 6.1 is close to cost-neutral. The savings come from stepping down to Terra or Luna for work that does not need a flagship, and the extra cost comes from stepping up to Astra. Reasoning tokens bill as output, so a task that thinks hard on Astra can cost well above this table. On a ChatGPT plan, the same task draws from your message allowance instead of tokens, which the table does not model.

Plan access or API key

Codex users reach these models in two ways, and the model list differs by route. With a ChatGPT sign-in, the plan decides which models you see, and usage is shared with ChatGPT Work and capped by five-hour windows. With an API key, you can use any model your key can access and pay token prices, but OpenAI's Codex pricing page says cloud features such as GitHub code review and Slack are not available on that route.

That split is why a Codex-branded model never mattered much to buyers. What mattered was the plan tier and the allowance. If you are choosing between routes, our Codex pricing page has the plan table and the arithmetic. For the contrast with a Claude-only agent, note that Claude Code uses Claude models alone, so its choice is among Fable, Opus, Sonnet and Haiku, described on Claude models.

Maker, modalities and context for the current choice

The full picture for the GPT-6 family is on GPT models.

Where it is served

Codex-branded models were API and Codex-only. We found none on any host in our directory. For the current GPT-6 models, OpenAI's own API is the source. We did not check any listed host's catalogue for GPT-6, so use each host's own model list, for example AWS Bedrock.

Which coding agents use them

OpenAI Codex CLI is the home of the family. It is Apache-2.0 open source, signs in with a ChatGPT plan or an API key, and takes a model through /model. GitHub Copilot's plans page still lists GPT-5.3-Codex among its models, and that list may change as the model retires on 1 April 2027. Cursor, Amp and OpenCode carry OpenAI models, and we did not check which of them still serve GPT-5.3-Codex.

Published benchmarks

OpenAI's model pages carry no benchmark tables for these models, and we found no independent scores that we could verify. Where the GPT-6 line appears on benchmarks, the only tables we found are in Anthropic's launch posts, covered on Claude Opus and GPT models. We repeat none of them here, because they describe GPT-6 Astra and GPT-5.6 Sol and not Codex-branded models.

Local hardware needs

None for any Codex model. They never had public weights. For a Codex-style coding agent on your own hardware, see best open source LLM for coding, where open-weights coding models and their model cards are covered.

Migration checklist

If your code still names a Codex model, work through this.

• Search config, CI files and scripts for codex model names.
• Replace each with the recommended model from OpenAI's deprecations table.
• Re-run your evaluation set, because tool behaviour and token use differ across generations.
• Check reasoning effort. GPT-6.1 Sol defaults to medium, and effort up to max is available.
• Re-estimate cost. Output tokens are cheaper on Sol 6.1 than on GPT-5.3-Codex, input slightly dearer.
• Watch the 272K prompt threshold, above which input is priced at 2x and output at 1.5x on API use.
• Recheck the deprecations page before you commit; this page is dated 31 October 2026.

Who should care

Our take

The Codex line was a short experiment in a separate coding branch, and OpenAI has folded it back into the main models. That is a useful signal. For most buyers the question is no longer "which Codex model" but "which GPT-6 model and which reasoning effort." Sol 6.1 is the natural default for agent work, Terra 6.1 for cheaper steps, and Luna for volume.

The practical risk is old names lingering. A script pinned to gpt-5.2-codex has been dead since July. We recommend grepping for them now. We have not tested any of these models; this page is a record of OpenAI's own documents. For the agent itself, see OpenAI Codex CLI, and for how it compares with Anthropic's, Claude Code vs Codex.

Questions we hear

Was GPT-5.3-Codex better than GPT-5.2-Codex? OpenAI describes it as "the most capable agentic coding model to date" on its page. We have no scores to compare.

Can I still call GPT-5.3-Codex? Until 1 April 2027, per the deprecations page, though it is already marked deprecated.

Is there a free Codex model? Codex's Free and Go plans mention GPT-6 Luna in the desktop app, subject to rollout. See Codex pricing.

Why did OpenAI stop making Codex-only models? We found no statement of the reason. The deprecations page simply points the old names at general GPT-5.6 and GPT-6 models, which suggests the general line now covers coding.

Do I lose features by moving off a Codex model? The old pages list Responses API support and reasoning tokens, and the GPT-6 pages list the same endpoints plus streaming, structured outputs, function calling, file search and web search on Astra. Check the "Supported tools" section on the model page you move to.

What about the Mini model? GPT-5.1-Codex Mini at $0.25 and $2 is gone, and OpenAI's mapping sends it to GPT-5.6 Terra. In the GPT-6 line, the cheap tiers are Terra 6.1 and Luna.