OpenAI’s Thibault Sottiaux announced on October 12, 2026 that GPT-6.1 Sol now defaults to a 500k context window in the Codex and Work environments, up from 272k — an increase of about 80%. The maximum is still 1M if you raise it by hand, and OpenAI says it will not charge long-context pricing when GPT-6.1 Sol usage is deducted from a subscription.
Key facts
| Fact | Detail |
|---|---|
| Question answered | GPT-6.1 Sol Defaults to 500k Context in Codex and Work |
| Short answer | OpenAI’s Thibault Sottiaux said on October 12, 2026 that GPT-6.1 Sol now defaults to a 500k context window in Codex and Work, up from 272k, or about 80% more. |
| Reading time | ~3 min |
| Last updated | 2026-10-12 |
What changed
- New default: 500k tokens, up from 272k, in Codex and Work.
- Manual control stays. You can still set the window yourself, up to a maximum of 1M.
- No long-context surcharge. For subscription usage deducted from a plan, OpenAI says it will not bill the long-context rate for this model.
- The stated goal is fewer compactions, so work that needs to keep a large amount of detail around finishes with fewer interruptions.
The announcement and the figures above are from ithome.com’s October 12 report.
Why the default matters more than the maximum
A maximum is a specification; a default is what your runs actually use. Almost nobody opens a settings panel before asking an agent to refactor a module, and every agent session that hits the limit gets compacted: earlier turns are summarised, details are dropped, and the model keeps working from a lossy version of the conversation.
Moving the default from 272k to 500k roughly doubles the room before that happens. For long sessions — a task list that has been running for an hour, or a review that keeps referring back to files read early on — that is the difference between carrying the original detail and re-reading it. Sottiaux’s framing is that the number was tested and is the sweet spot between capability and cost for most people, not that bigger is always better.
The honest caveat is that a bigger default does not make an agent better at using context. Every tool result and file dump still competes for the same space, and a 500k window filled with unfiltered output is not obviously better than a 272k window filled with the right 30k. What it does remove is the case where a careful, well-scoped session still gets compacted because the model started with a small budget.
What this means if the agent runs on your machine
Context length is a server-side setting, and it is worth saying plainly that it only applies to the model doing the thinking. If your agent runs locally on Windows — proposing steps, waiting for your approval and then touching your real files and commands — the local side still has to decide what to feed the model and what to keep out. A larger window makes that filtering job less desperate, not unnecessary, which is why approval-first desktop agents like MOOGH stay useful either way; it is a free download.
What to watch next
- Whether the higher default holds at peak hours or quietly reverts under load.
- Whether the no-long-context-surcharge policy survives the next pricing change.
- Whether other models in the same family get the same treatment, or this stays a GPT-6.1 Sol decision.
FAQ
- Do I need to change anything? No. The new default applies on its own in Codex and Work; you only touch the setting if you want something other than 500k.
- Can I lower it? Yes. The value is still manually configurable, both up to the 1M maximum and down if you prefer a smaller window.
- Does this change the model’s quality? No. A longer window changes how much of the session survives before compaction, not what the model is capable of on a given prompt.
Related reading: