OpenAI has introduced GPT‑5.1-Codex-Max, a brand new frontier agentic coding mannequin now obtainable in its Codex developer surroundings. The discharge marks a major step ahead in AI-assisted software program engineering, providing improved long-horizon reasoning, effectivity, and real-time interactive capabilities. GPT‑5.1-Codex-Max will now exchange GPT‑5.1-Codex because the default mannequin throughout Codex-integrated surfaces.

The brand new mannequin is designed to function a persistent, high-context software program growth agent, able to managing advanced refactors, debugging workflows, and project-scale duties throughout a number of context home windows.

It comes on the heels of Google releasing its powerful new Gemini 3 Pro model yesterday, but nonetheless outperforms or matches it on key coding benchmarks:

On SWE-Bench Verified, GPT‑5.1-Codex-Max achieved 77.9% accuracy at extra-high reasoning effort, edging previous Gemini 3 Professional’s 76.2%.

It additionally led on Terminal-Bench 2.0, with 58.1% accuracy versus Gemini’s 54.2%, and matched Gemini’s rating of two,439 on LiveCodeBench Professional, a aggressive coding Elo benchmark.

When measured in opposition to Gemini 3 Professional’s most superior configuration — its Deep Pondering mannequin — Codex-Max holds a slight edge in agentic coding benchmarks, as properly.

Efficiency Benchmarks: Incremental Positive aspects Throughout Key Duties

GPT‑5.1-Codex-Max demonstrates measurable enhancements over GPT‑5.1-Codex throughout a spread of ordinary software program engineering benchmarks.

On SWE-Lancer IC SWE, it achieved 79.9% accuracy, a major improve from GPT‑5.1-Codex’s 66.3%. In SWE-Bench Verified (n=500), it reached 77.9% accuracy at extra-high reasoning effort, outperforming GPT‑5.1-Codex’s 73.7%.

Efficiency on Terminal Bench 2.0 (n=89) confirmed extra modest enhancements, with GPT‑5.1-Codex-Max attaining 58.1% accuracy in comparison with 52.8% for GPT‑5.1-Codex.

All evaluations had been run with compaction and extra-high reasoning effort enabled.

These outcomes point out that the brand new mannequin affords a better ceiling on each benchmarked correctness and real-world usability beneath prolonged reasoning hundreds.

Technical Structure: Lengthy-Horizon Reasoning by way of Compaction

A serious architectural enchancment in GPT‑5.1-Codex-Max is its potential to purpose successfully over prolonged input-output periods utilizing a mechanism referred to as compaction.

This permits the mannequin to retain key contextual data whereas discarding irrelevant particulars because it nears its context window restrict — successfully permitting for steady work throughout hundreds of thousands of tokens with out efficiency degradation.

The mannequin has been internally noticed to finish duties lasting greater than 24 hours, together with multi-step refactors, test-driven iteration, and autonomous debugging.

Compaction additionally improves token effectivity. At medium reasoning effort, GPT‑5.1-Codex-Max used roughly 30% fewer pondering tokens than GPT‑5.1-Codex for comparable or higher accuracy, which has implications for each value and latency.

Platform Integration and Use Instances

GPT‑5.1-Codex-Max is at present obtainable throughout a number of Codex-based environments, which discuss with OpenAI’s personal built-in instruments and interfaces constructed particularly for code-focused AI brokers. These embrace:

• Codex CLI, OpenAI’s official command-line software (@openai/codex), the place GPT‑5.1-Codex-Max is already stay.

• IDE extensions, seemingly developed or maintained by OpenAI, although no particular third-party IDE integrations had been named.

• Interactive coding environments, corresponding to these used to show frontend simulation apps like CartPole or Snell’s Regulation Explorer.

• Inside code assessment tooling, utilized by OpenAI’s engineering groups.

For now, GPT‑5.1-Codex-Max just isn’t but obtainable by way of public API, although OpenAI states that is coming quickly. Customers who want to work with the mannequin in terminal environments right this moment can achieve this by putting in and utilizing the Codex CLI.

It isn’t at present confirmed whether or not or how the mannequin will combine into third-party IDEs except they’re constructed on prime of the CLI or future API.

The mannequin is able to interacting with stay instruments and simulations. Examples proven within the launch embrace:

• An interactive CartPole coverage gradient simulator, which visualizes reinforcement studying coaching and activations.

• A Snell’s Regulation optics explorer, supporting dynamic ray tracing throughout refractive indices.

These interfaces exemplify the mannequin’s potential to purpose in actual time whereas sustaining an interactive growth session — successfully bridging computation, visualization, and implementation inside a single loop.

Cybersecurity and Security Constraints

Whereas GPT‑5.1-Codex-Max doesn’t meet OpenAI’s “Excessive” functionality threshold for cybersecurity beneath its Preparedness Framework, it’s at present probably the most succesful cybersecurity mannequin OpenAI has deployed. It helps use instances corresponding to automated vulnerability detection and remediation, however with strict sandboxing and disabled community entry by default.

OpenAI stories no improve in scaled malicious use however has launched enhanced monitoring methods, together with exercise routing and disruption mechanisms for suspicious habits. Codex stays remoted to a neighborhood workspace except builders opt-in to broader entry, mitigating dangers like immediate injection from untrusted content material.

Deployment Context and Developer Utilization

GPT‑5.1-Codex-Max is at present obtainable to customers on ChatGPT Plus, Professional, Enterprise, Edu, and Enterprise plans. It is going to additionally change into the brand new default in Codex-based environments, changing GPT‑5.1-Codex, which was a extra general-purpose mannequin.

OpenAI states that 95% of its inner engineers use Codex weekly, and since adoption, these engineers have shipped ~70% extra pull requests on common — highlighting the software’s impression on inner growth velocity.

Regardless of its autonomy and persistence, OpenAI stresses that Codex-Max needs to be handled as a coding assistant, not a alternative for human assessment. The mannequin produces terminal logs, check citations, and power name outputs to assist transparency in generated code.

Outlook

GPT‑5.1-Codex-Max represents a major evolution in OpenAI’s technique towards agentic growth instruments, providing higher reasoning depth, token effectivity, and interactive capabilities throughout software program engineering duties. By extending its context administration and compaction methods, the mannequin is positioned to deal with duties on the scale of full repositories, slightly than particular person information or snippets.

With continued emphasis on agentic workflows, safe sandboxes, and real-world analysis metrics, Codex-Max units the stage for the following era of AI-assisted programming environments — whereas underscoring the significance of oversight in more and more autonomous methods.

Source link