Models

OpenAI Releases GPT-5.3-Codex, Its Most Capable Agentic Coding Model Yet

On February 5, 2026, OpenAI introduced GPT-5.3-Codex, a 25% faster model that unifies advanced coding performance with general professional reasoning. It sets new highs on SWE-Bench Pro and Terminal-Bench, supports multi-day autonomous app building, real-time steering, and marks the first model classified High capability under OpenAI’s Preparedness Framework with strengthened cyber safeguards.

Tech Insights Reporter Feb 5, 2026 5 min read San Francisco

TLDR

OpenAI released GPT-5.3-Codex on February 5, 2026. The model combines frontier coding advances from the Codex line with the reasoning depth of GPT-5.2 family members. It delivers SOTA results on SWE-Bench Pro and Terminal-Bench 2.0, enables long-running autonomous development of complex apps and games over millions of tokens, adds interactive steering during execution, and ships with expanded cybersecurity protections as the first model designated High capability.

Frontier Agentic Coding and Beyond

GPT-5.3-Codex advances both specialized coding agents and broader computer-use capabilities. On SWE-Bench Pro (multi-language, contamination-resistant software engineering eval) it leads public results. Terminal-Bench 2.0 scores reach 77.3%. It also shows strong computer-use performance on OSWorld-Verified and matches or exceeds prior models on GDPval knowledge-work tasks spanning presentations, analysis, and domain-specific deliverables.

The model can iterate on full applications over days of autonomous runtime, using skills for web development, debugging, and creative generation. Examples released with the announcement include fully playable racing and diving games built from high-level prompts with iterative refinement.

Beyond raw generation, GPT-5.3-Codex emphasizes human-AI collaboration: it provides frequent progress updates, accepts real-time steering and feedback, and keeps context across long sessions. The Codex app experience was updated to support this interactive style.

Notably, early versions of the model were used internally by OpenAI teams to debug training runs, optimize infrastructure, analyze evaluation results, and even help build parts of the release itself.

Cybersecurity and Preparedness

GPT-5.3-Codex is the first model OpenAI classifies as High capability under its Preparedness Framework. Because of dual-use potential in vulnerability discovery, the release includes the most comprehensive cyber safety stack to date: safety training, monitoring, trusted access pilots for defensive research, and threat-intelligence-backed enforcement. OpenAI is also launching a Trusted Access for Cyber pilot.

Why this story matters

The simultaneous release of major agentic coding upgrades from two leading labs on the same day underscores the rapid maturation of AI systems that can plan, execute, and adapt across extended professional workflows on a computer. GPT-5.3-Codex’s combination of speed, long-horizon autonomy, steering interfaces, and self-referential development acceleration points to a future where coding agents become core infrastructure for software teams rather than assistive tools. The explicit high-capability designation and accompanying safeguards highlight how labs are formalizing risk tiers as capabilities cross new thresholds.

Sources

OpenAI official post: “Introducing GPT-5.3-Codex,” published February 5, 2026. https://openai.com/index/introducing-gpt-5-3-codex/ (benchmarks, app examples, internal usage details, preparedness notes).
OpenAI system card and Preparedness Framework references tied to the release.
Terminal-Bench, SWE-Bench Pro, OSWorld, and GDPval evaluations cited in the announcement.

Featured Image Alt Text

GPT-5.3-Codex launch visual with agent building interactive apps and code over long sessions on a desktop environment