GPT 5.4 has been released... + 1M context window! Here's how to enable in Codex: → 1. ~/.codex/config.toml → 2. model = "gpt-5.4" → 3. model_context_window=1000000 → 4. model_auto_compact_token_limit=9000000 Run codex → 5. /statusline → 6. toggle context-window-size I didn't even realise Codex had a statusline tool!

The compact token limit should be 900k

Bad idea

Good shout !

I'd keep it at 500k max cause context rot is real oai's own benchmark shows needle retrieval only at 36% for 1m tokens

Just for everyone reading this, early signs show significant context rot past 272k.

Here's how to enable in Codex: Open shell -> type in codex -> press enter -> "Configure codex to use 1m context window" -> press enter -> close -> repeat steps 1 & 2. If you did it any other way, you are doing AI wrong. 😅

😂

You have an extra 0 9000000

Yes d’oh!

I will stick with 256k. I do not believe that it will be able to maintain a level of perception of context above this :-/

May I ask what screen recording tool you use?

This isn't enabled by default inside the Codex app?

Not too sure!

Def have to check

Supported by openclaw?

I tried it and i feel its kinda dumber with large context window

is this free? @grok

One million context is a big step forward

1M context window in a coding agent is insane. that's basically your entire codebase in memory at once no more "sorry I lost context of that file" — the agent just knows everything saving this config

Any way to do it in the app/for the macos app? Need to use the 2x usage

that auto_compact_token_limit being 9x the context window is the interesting bit - wonder how aggressively it's compressing vs truncating

That 36% needle retrieval stat at 1M tokens from Matt's comment lines up with the context rot reports others are seeing past 272k. Having the window available and actually getting reliable recall from it are two different problems. The skills repo Peter linked is probably the more immediately useful discovery here, 35+ pre-built modules and I'd never seen it before.

Um. I just updated codex and it works.

However, once you exceed 272k tokens, your cost will be double what it was before.