Reopen #19464 and support 1M context
What variant of Codex are you using?
App,CLI
What feature would you like to see?
Ever since GPT-5.5 was released, the context window has been stuck at only 272k. Any related discussion gets immediately closed by staff with the "openai" suffix.
At this point, 272k context is practically unusable in codex. I run into context compression failures more than once a day, and each thread averages five to six context compressions. Every compression noticeably degrades work capability and causes the model to forget things.
Why not give us back the 1M context? Even an option like a [1m] toggle like the one in claude code wouldn't work? Consume more credits if you have to. Why actively downgrade developers' experience?
If you're worried about compute capacity, don't enforce any actual limit on the server side—just let us flip this hidden option through an env variable so that those who need it can enable it. Would that really be so hard?
Additional information
_No response_
10 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
no dublicate.
I don't think this is a technical limitation—it's a cost issue. The actual API cost of Codex subscriptions likely far exceeds the subscription price, and supporting a 1M-token context window would only drive those costs much higher. At least for now, OpenAI probably isn't planning to absorb that expense for developers.
Right now, Codex is geared more toward localized feature improvements and is still far from being able to replace CC. As long as the context window doesn't grow, it won't be able to handle large-scale architectural tasks.
So I don't think developers need to keep pushing Codex to support a 1M-token context window. Once OpenAI is rich enough, they'll support it naturally.
They can set a higher price, and I'm willing to pay more for it.
The subscription price is already far below the actual API cost. Even charging developers a few hundred dollars more wouldn't be enough to cover the additional cost of supporting a 1M-token context window.
It was implied in the original issue that this would be implemented. It's honestly the only item stopping me from cancelling my Claude sub and going all in OpenAI.
Noticed the same drop — went from 1M to 272K mid-sprint.
Coincidentally, I've been working on a methodology for exactly this scenario: Deterministic Context Routing (DCR) — structured context management designed for short-window agent sessions. Module registry, layered knowledge base, session state with stop markers, confidence labels. The idea is to route only what's needed instead of dumping everything into one window.
There's a write-up here: [https://github.com/ai-erp-collab/deterministic-context-routing](url)
It helps with large, raw-data-heavy projects — the kind where you'd normally burn through context just navigating the workspace. ERP development, multi-module codebases, that sort of thing.
Might be useful if the window stays where it is.
It's not dropping midsprint. It will show 1M before you send the first prompt and then will update right after that or after the first response. Which probably means you configured codex to use 1M, but it can't and only updates the UI when you try.
Fable 5: 1M. K3: 1M. GLM-5.2: 1M. Grok 4.5: 500K. GPT-5.6 Sol API: 1.05M. In Codex: 272K.
OpenAI built a frontier brain and locked it in a toy cage. 272K isn’t context.
It’s a README with delusions of grandeur. Codex didn’t deploy Sol. It imprisoned it.
<img width="1200" height="1200" alt="Image" src="https://github.com/user-attachments/assets/24944441-918e-4232-8d8d-489ed2a86d85" />