Repository navigation
Codex: default vs selected variants show different effective context; support per-variant limits #12416
Closed as not planned
Description
Summary
Observed behavior suggests default Codex and selected variants can have different effective context limits.
Observation
For openai/gpt-5.3-codex:
- With default settings (no variant selected), effective context can exceed ~272k.
- When explicitly selecting variants like
high/medium, effective context appears capped around ~272k.
This makes default vs selected variants behave differently in practice.
Reference comment in #7705:
#7705 (comment)
Request
Please consider supporting per-variant limit configuration (or equivalent behavior) so users can express this explicitly.
Example config shape:
{
"provider": {
"openai": {
"models": {
"gpt-5.3-codex": {
"limit": { "context": 400000, "output": 128000 },
"variants": {
"high": {
"reasoningEffort": "high",
"limit": { "context": 272000, "output": 128000 }
}
}
}
}
}
}
}Even if internals normalize limits differently, allowing variant-level declaration (or rejecting unsupported keys with a clear error) would reduce confusion.
Reacted by piljoong-jeong and Oscar Serna
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Feb 6, 2026
This issue might be a duplicate of existing issues. Please check:
Feel free to ignore if none of these address your specific case.
HyunggyuJang commentedon Feb 7, 2026
Closing this as my report appears incorrect.
I traced this to model metadata in pi-mono:
https://github.com/badlogic/pi-mono/blob/f9161c4d4e0d001ebfbd3ba98f0b89adb253260a/packages/ai/src/models.generated.ts#L5099
It looks like ChatGPT subscription-backed Codex models are already constrained (roughly 272k input / 128k output) regardless of selected variant, so the default-vs-variant interpretation here was likely mistaken.
Sorry for the noise.