GPT-5.1-Codex-Max
OpenAI's long-horizon agentic coding model with compaction, trained for multi-hour software work.
Retired
GPT-5.1-Codex-Max appears to be discontinued or shut down — Original API model ID deprecated; the Codex Max name now refers to a reasoning-effort mode on newer models This page is kept as an archive record and is excluded from search engines.
About
GPT-5.1-Codex-Max is OpenAI's agentic coding model for long-running software work, announced November 19, 2025 and listed on the API in December. It is trained on real engineering work like pull requests, code review, and iterative debugging, and it can stay on a single task for hours. The headline feature is compaction: the model prunes its own session history as it goes so it can operate coherently across multiple context windows, millions of tokens in total.
OpenAI reports 76.5% on SWE-bench Verified at high reasoning effort and 77.9% at the new Extra High setting, plus 58.1% on Terminal-Bench 2.0, both ahead of the base GPT-5.1-Codex. It was the default model in Codex surfaces at launch and the first OpenAI model to work smoothly in Windows environments. Note the naming trap: in current OpenAI documentation, "Max" is a reasoning-effort mode rather than this standalone model, and the original model ID is marked deprecated in the API catalog.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SWE-bench Verified: 76.5% (high) / 77.9% (xhigh), per OpenAI
- Terminal-Bench 2.0: 58.1%, per OpenAI
- Reported ~30% fewer thinking tokens than GPT-5.1-Codex at equal quality
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (verified tier, metrics-v1). Evidence: widely used as the default Codex model at launch with the original API model ID since deprecated; compaction technique carried into newer Codex models. Closed weights. Launch facts corroborated across multiple outlets; docked for OpenAI's self-reported benchmark figures and the 'Max' name now referring to a reasoning-effort mode rather than this model.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Adoption 6: default model in Codex surfaces at launch (Nov 2025); API model ID since marked deprecated, successor 5.2-Codex shipped Dec 2025
- · Maintenance 6: announced Nov 19, 2025, API-listed Dec 4, 2025; superseded by newer Codex models but the compaction technique carried forward
- · Openness 3: proprietary, no weights, API/ChatGPT-only
- · Manual 7: launch facts (compaction over multiple context windows; SWE-bench Verified 76.5%/77.9%; Terminal-Bench 2.0 58.1%; $1.25/$10 pricing) corroborated across OpenAI-adjacent reporting, Simon Willison, and OpenRouter; docked because OpenAI's benchmark figures are self-reported and current 'Max' docs now describe a mode, not this model