Thibault Sottiaux (Tibo), who leads OpenAI's Codex, announced late on October 11 on X that GPT-6.1 Sol's default context length in Codex and ChatGPT Work is going from 272K tokens to 500K tokens, an increase he put at roughly 80%.
It is day 7 of the Codex team's "28 days of updates" program. Announced on October 4, the program's rule is that every day the team either ships an improvement most users will benefit from or resets usage limits for everyone.
What changed
Sottiaux gave just one reason: fewer context compactions. During long tasks, Codex's agent keeps stuffing file contents, command output and tool-call records into its context. Once the window fills up, the system has to compress older material to keep going. The cost is lost detail, and the agent may forget a function signature or an error message it read earlier.
"It was carefully tested and is the new sweet spot for most of our users."
The new default can be raised or lowered manually. The ceiling he cited in the post is 1 million, while OpenAI's model page lists the full GPT-6.1 Sol window at 1.05 million tokens, with a maximum output of 128K. He also added "We will go higher in the future", without giving a timeline.
Some reports say the change carries no extra charge for subscribers. Sottiaux's post itself says nothing about pricing, and whether the larger default window will burn through weekly usage faster has not been addressed publicly by OpenAI; treat any follow-up usage guidance from the company as authoritative.
Where the 272K figure comes from
The old 272K default wasn't arbitrary. OpenAI's model page states that when an API request to GPT-6.1 Sol has more than 272K input tokens, the entire input and cached input are billed at 2x and output at 1.5x. At the standard rate of $2 per million input tokens and $10 per million output tokens, that works out roughly to $4 for input and $15 for output once the line is crossed.
The default and the surcharge threshold coincide exactly, and some analysts have inferred from this that the old setting was shielding users from long-context costs. Sottiaux has not confirmed that. Subscribers pay through usage allowances rather than per token, and outsiders cannot see how the two schemes convert into each other.
Connecting the recent dots
Codex's context management has been shifting for the past two months:
- In early September, several pull requests landed in the openai/codex repository. They move away from generating a summary when a session fills up, and instead open a new window and pair it with a history-retrieval tool so the model can look through old records itself;
- On September 30, a user filed a GitHub issue saying that in the Codex desktop app, starting a new thread and replying with just "OK" brought GPT-6.1 Sol's context usage to about 208K / 258K, compared with about 40K for GPT-6 Sol in the same scenario. The reporter said they couldn't tell whether the extra tokens were actually sent to the model or whether the display was wrong. The issue was marked completed and closed on October 8, with no reply from an OpenAI employee on the page;
- On October 11, the default window went to 500K.
Taken together, Codex is reworking compaction on one hand and enlarging the window on the other, and both point at the same symptom: agents "forgetting things" on long tasks. Whether this change is directly related to that issue report, the company hasn't said. If new threads still begin by consuming more than 200K tokens, the old 258K of usable space would run out almost as soon as work started, and at 500K there is considerably more headroom.
For developers in China, Codex requires a paid ChatGPT account. This change affects only GPT-6.1 Sol in the Codex and Work entry points. Projects calling the API directly are unaffected: the requester still decides the context length, and the pricing threshold is unchanged.
Sources: Thibault Sottiaux's social media announcement, the OpenAI GPT-6.1 Sol model page, CocoLoop, FourWeekMBA; the openai/codex repository issue report was used to check the new-thread context usage figures, and the model page to check the window ceiling and long-context pricing multipliers.