October 10, 2026
Bigger context doesn't mean smarter: in chess, an 800k model lost to a 300k version
In chess, a model with an 800,000-token window plays worse than with a 300,000-token window, confirmed on Oct 10, 2026 by the post's author @petergostev. This is one test, not a benchmark: the post has neither games nor ratings.
Peter Gostev
@petergostev
I can confirm it at least for chess: the model with an 800,000-token context plays worse than with a 300,000-token context.
@rohit3a Folks. Why do you think we picked 300k for context length for Codex. We ran the numbers and it’s better. We can run it at 1M too but it would just cost more of the usage. Good defaults are important.
· 19.1K views
Codex runs on a 300,000-token window, even though a 1 million token window is available. @thsottiaux's answer to the question about this choice: we ran the numbers, and this works better. The 1 million window can be turned on, but it eats more of your usage limit. The takeaway for practitioners is simple: window size is not the same as quality. A 1 million token window costs usage limit, and the extra tokens did not help in chess.
For now Codex defaults to a 300,000-token window, and 1 million stays an option that costs extra usage limit.