GPT-6.1 Sol now defaults to 500K context in Codex! But note: GPT-6 Astra still defaults to 272K (at least as of writing, Oct 13).

Recently, OpenAI’s Thibault Sottiaux announced the increase for GPT-6.1 Sol in Codex and Work. Sol went from 272K to 500K, roughly 80% more room. Tibo says they tested this size carefully to reduce compactions while keeping more context and nuance through long tasks.

That’s useful when a task spans a lot of code, rejected approaches, and constraints.

Both Sol and Astra support up to 1M context. Those maximums are separate from the defaults Codex chooses.

Leave the overrides unset

I used to configure model_context_window explicitly. Now that Sol and Astra have different defaults, that global setting gets in the way: it applies the same window to whichever model I select.

I’d leave it unset unless a task calls for a deliberate override. Codex’s sample configuration says an omitted setting uses the model or preset defaults.

If you’ve added this to ~/.codex/config.toml, you might want to remove it to follow the defaults:

model_context_window = 500000

Pinning 500K would override Astra’s 272K default. If a request grows from 272K to 500K input tokens, that’s around 2× the input-token usage 😱

I’d also leave model_auto_compact_token_limit unset. The configuration reference says that uses the model’s default compaction threshold. A low custom threshold can still trigger an early compact even with a larger window.

I suspect Astra will get a revised higher default too. That’s my guess.