[scottndecker]

Still 256K input tokens. So disappointing (predictable, but disappointing).

[coder543]

https://platform.openai.com/docs/models/gpt-5.2

400k, not 256k.

[nathants]

400 - 128 = 272. Codex cli source.

[coder543]

If you want to be able to generate up to 128k tokens in one go successfully, then yes, that math checks out.

[htrp]

much harder to train longer context inputs