GPT-5.1-Codex-Max Release Notes
OpenAI has introduced GPT-5.1-Codex-Max, a frontier agentic coding model designed for long-running, complex software engineering tasks. This model is built on an updated foundational reasoning model trained on agentic tasks across research, math, and software engineering, enabling it to act as a more reliable coding partner.
Long-Horizon Coding via Compaction
GPT-5.1-Codex-Max introduces a native process called "compaction," which allows the model to operate across multiple context windows. By pruning history while preserving critical context over long horizons, the model can coherently work over millions of tokens in a single task.
This capability enables:
- Project-scale refactors: The model can handle large-scale changes across a codebase without losing progress.
- Deep debugging sessions: Sustained reasoning over complex bugs.
- Multi-hour agent loops: In internal evaluations, the model has worked independently on tasks for more than 24 hours, persistently iterating on implementations and fixing test failures.
In Codex applications, the model automatically compacts its session as it approaches the context window limit, providing a fresh window to continue the task until completion.
Technical Performance and Token Efficiency
GPT-5.1-Codex-Max is more token-efficient and faster than its predecessors. On the SWE-bench Verified benchmark, the model achieves better performance than GPT-5.1-Codex using "medium" reasoning effort while consuming 30% fewer thinking tokens.
Reasoning Effort Levels
- Medium: Recommended as the daily driver for most tasks.
- Extra High (xhigh): A new setting for non-latency-sensitive tasks that allows the model to think for longer periods to achieve a higher quality answer.
Benchmark Results
| Benchmark | GPT-5.1-Codex (high) | GPT-5.1-Codex-Max (xhigh) |
|---|---|---|
| SWE-bench Verified (n=500) | 73.7% | 77.9% |
| SWE-Lancer IC SWE | 66.3% | 79.9% |
| Terminal-Bench 2.0 | 52.8% | 58.1% |
Agentic Capabilities and Environment Support
GPT-5.1-Codex-Max was trained on real-world software engineering tasks, including PR creation, frontend coding, and code review. It is the first model in the series trained to operate specifically in Windows environments and includes specialized training for better collaboration within the Codex CLI.
Security and Safety Frameworks
While GPT-5.1-Codex-Max is the most capable cybersecurity model OpenAI has deployed to date, it does not reach "High" capability under OpenAI's Preparedness Framework. To mitigate risks, OpenAI has implemented the following:
- Secure Sandboxing: Codex runs in a secure sandbox by default, restricting file writes to the workspace and disabling network access unless explicitly enabled by the developer.
- Cybersecurity Monitoring: Dedicated monitoring is used to detect and disrupt malicious activity, with suspicious activity routed through policy monitoring systems.
- Human-in-the-Loop: Because the model can perform long-running tasks, OpenAI recommends that developers review all agent work—supported by terminal logs and tool call citations—before deploying to production.
Availability and Integration
GPT-5.1-Codex-Max is now the default model in Codex surfaces and is available for users on ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. It is specifically recommended for agentic coding tasks in Codex or similar environments rather than general-purpose use.
API access for developers using the Codex CLI via API key is expected to be available soon.