Codex context window cut to 272K: why that number

Codex context window size fell from 372K to 272K

Codex is the coding agent from OpenAI that reads a repository, edits files and runs commands for you. With this change, the three GPT-5.6 models inside it lost 100,000 tokens of configured window each. Both the raw setting and the part a session can really use went down.

The cut came as a one-file GitHub change with no notice

PR #33972 carries the edit. Titled "Backport refreshed bundled model metadata to 0.144," it was merged into the Codex 0.144 release branch on July 18, 2026.

OpenAI posted nothing on its blog or in release notes, so the change first surfaced among developers who read the diff. Only one file changes: models.json, where the entries for GPT-5.6 Sol, Terra and Luna get both context_window and max_context_window rewritten from 372,000 to 272,000.

The 272,000 figure is not new for Codex. In the same file, GPT-5.4 and GPT-5.2 already carry a 272,000-token window, so the GPT-5.6 models are back in line with the generations before them.

Codex's 5% headroom puts the usable limit near 258K

Codex does not let a session fill the whole configured window. Issue #32486 notes that the model metadata applies a 95% effective-context multiplier, so 5% of the window is held back.

Before the cut, that gave an effective window of about 353,400 tokens. Applying the same 95% to 272,000 gives about 258,400, so the room a session can really use shrank by roughly 95,000 tokens.

Past 272K input, Codex usage limits run out faster

The new number sits exactly on a pricing line. GPT-5.6 Sol moves to a higher rate once input passes 272K tokens, and the old default let sessions cross that point without warning.

The whole request is billed at 2x input past 272K

OpenAI's model page for GPT-5.6 Sol says that a prompt with more than 272K input tokens is billed at 2x the input rate and 1.5x the output rate. The higher rate covers the full request, not only the tokens above the line, so one prompt that goes a little over costs about twice as much on the input side.

On an API key, the difference shows up on the bill. For people signed in with a ChatGPT plan, Issue #32486 calls the same line a "higher-usage" threshold: each request past it uses up the allowance faster.

The old 372K default could run 81,400 tokens past 272K

Issue #32486 lays out the arithmetic. With a usable window of about 353,400 tokens, a default session could run about 81,400 tokens into the higher band without the user ever choosing a larger context. The report adds that subagents, which Sol can launch to check its own work, add context the user did not ask for.

The workaround it offered was to set model_auto_compact_token_limit = 270000 so that Codex compacts before the line. After the cut, the default already does that: 258,400 usable tokens sit 13,600 below 272,000.

A smaller Codex token limit means earlier compaction

For small fixes, 258,400 usable tokens leaves plenty of room. The difference shows up in long sessions on large repositories, and in work that needs a very large context in one go.

GPT-5.6 context window: API vs Codex (tokens)

Bar height is relative to the API's 1,050,000 = 100. Orange marks the current Codex setting.

1,050,000GPT-5.6 Sol via the API

372,000Codex before July 18

272,000Codex now

0500,0001,000,000

Codex compaction kicks in sooner on large repositories

Compaction is what Codex does when the window fills: it replaces older turns with a summary to free up room. With 95,000 fewer usable tokens, that point arrives earlier in a session.

Each pass has a cost. Writing the summary takes model time, files read earlier may have to be read again, and details from early in the session survive only as far as the summary kept them. Work that loads many files at once, or runs for hours on one task, feels this first.

The GPT-5.6 context window in the API is still 1.05M

The 272K cap is a Codex setting, not a limit of the model. OpenAI's page for GPT-5.6 Sol still lists a 1,050,000-token context window with 128,000 max output tokens.

Codex itself can go higher too. PR #39102, merged on August 17, 2026 and shipped in Codex CLI 0.149.0, lets you override the window for the three GPT-5.6 models with model_context_window in config.toml, up to 872,000 tokens. The default is still 272,000, and a request over 272K input is billed at the higher rate whichever route you take.

Keep Codex sessions small while the cap stays at 272K

The limit is lower, but it now keeps default sessions out of the higher-priced band. A few habits help you get more work out of each session.

Scoping comes first, because every file Codex reads stays in the window until compaction. For measurements of how much an agent loads before you even type, see Claude Code Token Usage: 33K vs OpenCode's 7K.

Primary sources such as Codex PRs and issues are easier for an AI to read when you convert the page to Markdown first. The AI then gets the headings, tables and code blocks as structure instead of raw HTML.