What's this about?
Tibo Sottiaux documented a frequently-requested change: GPT-5.6 Sol supports a context window of about 1,050,000 tokens, but Codex sets a lower limit by default. You can raise it in `~/.codex/config.toml` by setting `model_context_window = 1000000` and `model_auto_compact_token_limit = 900000`. His caveat matters more than the setting: Codex's default was tuned for performance and cost, not chosen arbitrarily. Baoyu declined to change his, reasoning that with good compaction a shorter context performs better and costs less.
What exactly is a context window, and why not always maximize it?
The context window is how much text the model can hold at once — your code, the tool output, the conversation so far. When it fills, older material gets summarized away. Bigger sounds strictly better, but it isn't. Attention cost grows faster than linearly, so long contexts are slower and more expensive. And model quality degrades over very long contexts — the "lost in the middle" effect, where material buried in a huge context gets less effective attention than the same material in a compact one. A well-compacted 200k context often beats a sprawling 900k one.
What's the community saying?
- The demand is real even if the advice is "don't." Sottiaux documented it explicitly because it's a common ask, while noting the default is deliberate — an unusual "here's how, but consider not" post.
- Baoyu's counterargument is the useful one. Trusting the vendor's tuning and investing in good compaction instead is the more considered position, and it aligns with the same week's Pi compaction discussion.
- It reframes context as an economic choice. Because you're billed per token and long contexts slow generation, the window size is a cost-performance dial rather than a capability setting.