Codex 28 · Theme · #60

Per-model context window settings

7people describe this

7 on X

2 problems · 2 ideas · 3 mixed · first reported 11 Oct 2026

X replies to the promise post are counted apart from Rantr votes. “This happened to me too” adds your browser to this theme on Rantr, once.

Where OpenAI stands

Not addressed yet

No recorded announcement has answered this yet.

In their words

What people said

Idea
A "fix" for this would be to decrease the context window size in your config. 10/11 they upped default from 272k to 500k. Enter your codex config.toml, and add model_context_window = 272k model_auto_compact_token_limit = 244.8k ~60% less usage from what I've noticed.

@LeiXing24 on X ↗ · 18 Oct 2026
Mixed
as of oct 11, sol in codex defaults to a 500k context, up from 272k. a session can now grow to ~450k before it compacts, instead of ~245k. openai's pricing page says holding more context means more usage per message. to go back to 272k, set this at the top of…

@joshuaday on X ↗ · 17 Oct 2026
Mixed
Yeah I'd also like to understand this. It's great that we get up to 500k for the standard price but sometimes I still use the 1M override for long detail-sensitive threads.

@skaldebane on X ↗ · 13 Oct 2026
Problem
Oh so maybe this is the reason Sol 6.1 caused regressions in my massive projects? In that case I have changed my toml config to 1 million token window, so next time there is a reset I will give Sol 6.1 another chance and maybe with 1 million tokens it will do the work I need done…

@DeathStarRobot on X ↗ · 13 Oct 2026
Idea
here's my temp fix... add the override back... model_context_window = 500000 model_auto_compact_token_limit = 450000

@btgille on X ↗ · 12 Oct 2026
Problem
Dude, it's totally useless when you cannot set per-model context limit in codex. Setting 6.1-Sol to 500k makes every other model more expensive.

@salmonpod on X ↗ · 12 Oct 2026
Mixed
Tibo, I already set 500K through the config override. What’s new beyond making it the default? If that’s the whole change, it’s an overdue defaults change, not a new context capability.

@huss60 on X ↗ · 11 Oct 2026

Your experience

Had this too? Say what happened.

JEV reads your words as you type and shows which theme they belong to. It may place them in another theme if that fits better.

What happened with “Per-model context window settings”?

0 / 1200

Tell us a little more — at least 20 characters.

Your words stay yours.

AI reads your draft as you type. Only what you publish is public. Privacy

Not sure where to start?

Think of the last task you tried. What were you doing, what happened, and what would have helped? You can also say you haven't tried a change yet.

01
What you said

Read as you type: a win, a problem or an idea, and the words that carry it.
Ready when you are
Who else had it

The theme your words join, and how many people on Rantr and X described it.
Where OpenAI stands

Shipped, reset, or not addressed yet, from the recorded announcements.
What to do now

A step the announcements make possible for you, when there is one.

Status

Where OpenAI stands

“Shipped” means an announcement was linked to the theme by AI, not that the problem is fixed for everyone. “Not addressed yet” means no recorded announcement was linked; it is not proof that nothing changed.

No recorded announcement has been linked to this theme yet.

Corrections

Changes to this theme

14 Oct 2026, 06:00 UTC

Renamed from “Configurable context window size” to “Per-model context window settings” as more people described it.
13 Oct 2026, 05:43 UTC

“Per-model context limits missing” was merged into this theme: JEV judged them the same problem.

Follow this theme by email

We send a confirmation link first; nothing else is sent until you confirm. Then at most one email a day, only when the theme changes.

X thread: 2769 replies to the promise post collected (4 Oct 2026 – 26 Oct 2026): 2555 placed in a theme and 214 set aside as off-topic or unsafe. X replies are found by search and each one is checked against its public post on X. Last read: 26 Oct 2026, 23:46 UTC. Not every reply is read.

← Every theme