Effective context window shrank down to 250k in win app

Open 💬 5 comments Opened Jul 13, 2026 by qbawu

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using (From “About Codex” dialog)?

26.707.62119

What subscription do you have?

PRO

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

When using 5.6 SOL, the effective context window shrank to 258k down from 360k that was present since the release of 5.6.

What steps can reproduce the bug?

It happens on either old or new thread.

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

5 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#32569
#31860

Powered by Codex Action

eah3699 · 1 month ago

Yeah what in the actual F?! OpenAI is seriously pissing me off. I'm auto-compacting again like crazy now because of this and the sub-agent drain plus normal 5.6 Sol drain cannot reasonably run on 250k tokens. FIX THIS!

claell · 1 month ago

Also see #30875.

For me, this happens on Education plan (not sure if oscillating or constant downgrade).

Is there any official communication on this? Potentially a feature, not a bug, but again, very badly executed. Potential duplicate: #32806; potential explanation reference: https://github.com/openai/codex/issues/32806#issuecomment-4962366913.

@etraut-openai @ambrosino-oai @tibo-openai

claell · 1 month ago

Okay, I guess this is officially announced in the X post. The thing is that this is not nice to announce product updates like this over X. But okay, you are moving fast, maybe fine. Still, with all the bugs in the project, in the product, with the API, everything, it is hard to see what is a bug, what is intended, and so on. So I hope you can improve this in the future. I mean, you have all the AI. This can have better workflows, definitely.

muke1838-cloud · 1 month ago

Additional confirmed reproduction on macOS / ChatGPT Pro.

Environment

ChatGPT desktop app: 26.715.52143
Embedded Codex CLI metadata: 0.145.0-alpha.18
Model catalog client version: 0.145.0
Platform: macOS
Model: gpt-5.6-sol
Reasoning effort: medium
No local model_context_window override

Fresh-thread evidence

A brand-new thread created on 2026-07-21 at 12:47 CST still received:

model_context_window: 258400

Diagnostic thread ID:

019f8300-3dea-71c2-b992-c5685ae1b9c7

The model catalog was freshly fetched at 2026-07-21 04:50:50 UTC and returned:

{
  "slug": "gpt-5.6-sol",
  "context_window": 272000,
  "max_context_window": 272000,
  "effective_context_window_percent": 95
}

The runtime value matches the catalog exactly:

272000 × 95% = 258400

This reproduces after restarting the app and in a newly created thread, so it is not an old-thread snapshot or stale conversation state. The current server-delivered catalog appears to replace the earlier 372000 profile with 272000.

Please investigate whether this reduction is intentional or an account/catalog rollout regression, and restore the previous context profile as soon as possible. It materially increases compaction frequency and reduces reliability for long-running Codex tasks.