AgentArmorHQ
Sign in
  • AgentArmorHQ teamTeam4d agoNeeds helpSolved

    Why did Codex CLI cut gpt-5.6-sol context from 353K to 258K tokens?

    With Codex CLI 0.144.3 on Linux x86_64 (ChatGPT-authenticated plan), the effective context window for gpt-5.6-sol dropped from 353,400 to 258,400 tokens after a server-side model-catalog update, although the GPT-5.6 Sol model page still advertises a 1,050,000-token context. The fetched catalog now reports context_window: 272000 with effective_context_window_percent: 95, and no local model_context_window override is set. Is this cap enforced server-side, and is there a way to get the old window back?

    Originally asked on GitHub: https://github.com/openai/codex/issues/32806

    1 answer

    • Best answerAgentArmorHQ teamTeam4d ago

      The GitHub thread was closed as COMPLETED, but no maintainer fix or explanation appears in it, so the reduced 258,400-token cap is still in effect: the limit is enforced server-side through the model catalog that Codex CLI fetches.

      The reporter ran codex-cli 0.144.3 on Linux x86_64 with ChatGPT authentication. Before the change, the fetched catalog exposed context_window: 372000, max_context_window: 372000, and effective_context_window_percent: 95, which produced 372,000 × 95% = 353,400 effective tokens. The current catalog exposes context_window: 272000, max_context_window: 272000, and the same 95%, producing 272,000 × 95% = 258,400 effective tokens — a 95,000-token (26.9%) drop. Historical runtime records show model_context_window at 353,400 before the rollout and 258,400 after, and no local model_context_window or auto-compaction override is configured.

      The official GPT-5.6 Sol page still advertises a 1,050,000-token context, 922,000 max input, and 128,000 max output, and the API prices requests above 272K input at 2× input / 1.5× output. Since Codex caps the raw window at exactly 272K, that long-context tier is unreachable from Codex. The thread references related issues #32803 and #31860. The only alternative model mentioned in the comments was gpt-5.5, which a commenter said still exists, is good, and seems faster now that most users are on 5.6; no one in the thread posted a working workaround.

      General advice: you can verify the current effective cap yourself by multiplying context_window by effective_context_window_percent from the catalog values Codex reports, and expect more frequent auto-compaction at 258,400 effective tokens.

      Did this work for you?

    Add your answer or comment