Codex keeps context at 300k on purpose: Tibo says 1M would just burn more of your usage

#What changed

Nothing shipped, but Tibo (Thibault Sottiaux, who leads Codex at OpenAI) explained a default people keep asking about. Replying on X to a thread about trimming Claude's auto-compaction to save usage, he said OpenAI picked 300k tokens as Codex's context length after running the numbers, because it worked better. Codex could run at 1M, he said, but that would mostly eat more of your plan's usage. His words: good defaults are important.

#Who is affected

Anyone using Codex on a ChatGPT plan, especially heavy users who hit weekly limits, and anyone who has overridden the context window or compaction settings in config.toml.

#What to do now

• Leave the context window at the default unless a specific task genuinely needs more history in view. A bigger window means each turn can carry more tokens, which spends more usage.
• If you changed it, check your ~/.codex/config.toml. The keys are model_context_window (context tokens available to the model) and model_auto_compact_token_limit (the token count that triggers automatic history compaction; unset uses the model's default). Both are listed in the Codex config reference.
• For long sessions, start a fresh thread per task instead of pushing one thread toward a huge context.

#What is not confirmed

OpenAI has not published the numbers behind the 300k choice; this is Tibo's statement on X. The thread he replied to (@rohit3a) claimed setting Claude Code's autocompact to 400k would have saved 29% of weekly usage; that is one user's field report about Claude, not independently verified, and we have not measured a similar effect in Codex.

Sources: Tibo on X, @rohit3a on X, Codex config reference.