GPT-5.1-Codex-Max

2025-11-19 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches GPT-5.1-Codex-Max, its new agentic development model, trained to work across multiple context windows through compaction, for tasks lasting several hours.

What changes

First model natively trained to operate across multiple context windows through compaction: it prunes its history while keeping the essentials and works over millions of tokens.
Observed working on its own for more than 24 hours on internal tasks.
About 30% fewer thinking tokens than GPT-5.1-Codex at medium effort for better results, and a new “xhigh” effort.
First model trained for Windows environments.
Replaces GPT-5.1-Codex as the default model in Codex.

### Specifications - 400,000-token context (compaction beyond that), 128,000-token output, text and image input, text output, cutoff September 30, 2024. - Available in Codex (CLI, IDE, cloud, code review) for ChatGPT Plus, Pro, Business, Edu and Enterprise. - Responses API (gpt-5.1-codex-max) on December 4, 2025. - Does not exceed the “High” level in cybersecurity, but is the vendor's most capable model in this area at this date. ### Pricing Same pricing as GPT-5.1-Codex. | | Input | Cached input | Output | | --- | --- | --- | --- | | gpt-5.1-codex-max | 1.25 | 0.125 | 10 | ### Benchmarks Scores published by the vendor, with compaction and very high effort (xhigh), versus GPT-5.1-Codex (high). | Benchmark | GPT-5.1-Codex-Max | GPT-5.1-Codex | | --- | --- | --- | | SWE-bench Verified (500 tasks) | 77.9% | 73.7% | | SWE-Lancer IC SWE | 79.9% | 66.3% | | Terminal-Bench 2.0 | 58.1% | 52.8% | ### Key takeaways The model targets project refactors and agent loops lasting several hours. The vendor states that 95% of its engineers use Codex every week and ship about 70% more pull requests.

Links

Tags : LLM, OpenAI, Codex, GPT-5.1, Code, Agent