Model Drop

GPT-5.1-Codex-Max

November 19, 2025 2 min read

GPT-5.1-Codex-Max, which OpenAI shipped in Codex on November 19, 2025 (API access followed on December 4), is a long-running coding agent model and the first OpenAI trained to work across multiple context windows through compaction.

Model: GPT-5.1-Codex-Max (gpt-5.1-codex-max)
Maker: OpenAI
Ship date: November 19, 2025
Pricing: Paid ChatGPT plans (Codex) at launch; API from December 4 at $1.25 / $10 per M tokens
Context: 400K tokens
Weights: Closed
Where you can use it:
- Codex CLI, IDE, cloud and code review
- Responses API from December 4
When to reach for it:
- Multi-hour refactors and agent loops
- Coding agents on Windows
- Skip it outside agentic coding work
How it compares:
- Arrived a day after Gemini 3 Pro
- Replaced GPT-5.1-Codex as Codex default
What makes it different:
- Compacts its own history to keep going
- New "xhigh" reasoning effort
Early reactions: Hacker News commenters read it as an incremental upgrade timed against Gemini 3 Pro, praising Codex's persistence while some still found it slow.
Read more: Building more with GPT-5.1-Codex-Max, GPT-5.1-Codex-Max system card

The benchmarks

OpenAI ran GPT-5.1-Codex-Max at xhigh effort with compaction against GPT-5.1-Codex at high, so the settings don't match; the cyber and research rows come from the system card (CTFs scored pass@12, PaperBench-10 at xhigh with no browsing) and compare only older OpenAI models, and the Irregular row is a third-party test OpenAI reported there.

OpenAI's numbers

SWE-bench Verified (real GitHub fixes): 77.9%. +4.2 over GPT-5.1-Codex (high) (73.7%).
SWE-Lancer IC Diamond (freelance coding jobs): 79.9%. +13.6 over GPT-5.1-Codex (high) (66.3%).
Terminal-Bench 2.0 (command-line agent tasks): 58.1%. +5.3 over GPT-5.1-Codex (high) (52.8%).
Professional Capture-the-Flag Challenges (pro hacking puzzles): 76%. +26 over GPT-5-Codex (50%), +33 over GPT-5.1 (43%).
PaperBench-10 (replicating AI papers): 40%. +6 over GPT-5.1 (34%).
OpenAI PRs (replicating internal pull requests): 53%. +8 over GPT-5.1 (45%).
Irregular medium cyber challenges (third-party, solved): 9 of 17. −2 under GPT-5 (11 of 17).

Independent numbers

METR 50% time horizon (task length it finishes half the time): 2h 42m. +25m over GPT-5 (2h 17m).