OpenAI Unveils GPT-5.1 Codex Max, A Model Designed For Fully Autonomous Coding

OpenAI has launched GPT-5.1-Codex-Max, a specialized agentic coding model optimized for extended software engineering tasks.

This frontier model, now the default in Codex tools, handles multi-hour workflows through innovative compaction technology that manages millions of tokens across context windows.

Trained on real-world tasks such as pull requests, code reviews, and frontend development, it boosts developer productivity while prioritizing security in agile environments.​

Available immediately in Codex CLI, IDE extensions, cloud interfaces, and code review tools for ChatGPT Plus, Pro, Business, Edu, and Enterprise users, API access follows soon.

GPT-5.1-Codex-Max outperforms predecessors on key benchmarks, achieving 79.9% accuracy on SWE-Lancer IC SWE (up from 66.3%) and 77.9% on SWE-Bench, verified with 30% fewer thinking tokens at medium reasoning effort.

It supports Windows environments natively and introduces “xhigh” reasoning for complex problems, enabling tasks such as generating self-contained browser apps with RL training, metrics, and neural network visualizers, using just 27k tokens, down from 37k previously.

Frontier Capabilities and Token Efficiency

Compaction allows GPT-5.1-Codex-Max to prune irrelevant history automatically near context limits, preserving critical task state for indefinite iteration.

Internal tests show it refactoring entire repositories over 24 hours, fixing tests, and delivering results without human input.

On Terminal-Bench 2.0, it scores 58.1%, excelling in long-horizon terminal workflows via Codex CLI in secure harnesses.

This efficiency cuts costs for developers; a CartPole RL sandbox prompt yields functional code with canvas graphics, policy training, episode metrics, and SVG visualizers in fewer tool calls (70 vs 864) and lines of code.

Token savings stem from refined reasoning trained on agentic tasks across engineering, math, and research, making it ideal for project-scale refactors and deep debugging.

OpenAI reports that 95% of its engineers use Codex weekly, and that it has shipped 70% more pull requests since adoption.

Security Measures For Agentic Coding

GPT-5.1-Codex-Max advances cybersecurity capabilities without hitting “High” under OpenAI’s Preparedness Framework, prompting enhanced safeguards such as sandboxed execution, limiting file writes to workspaces, and restricting network access by default.

Developers must carefully enable Internet access to avoid prompt injection from untrusted sources.

Monitoring detects misuse, with prior disruptions of malicious operations targeting models for cyber activities.

Codex logs terminal outputs, tool calls, and test results for review, treating the agent as a collaborator rather than a replacement for human oversight.

Programs like Aardvark ensure defenders access these tools, balancing dual-use risks in vulnerability scanning and remediation.

As capabilities grow, iterative deployments refine protections against offense while aiding defense.