GPT-5.3-Codex is here — OpenAI's most capable agentic coding model

OpenAI launched GPT-5.3-Codex on Feb 5, 2026. It's their most capable agentic coding model so far — and the first one that was instrumental in creating itself. The team used early versions to debug training, manage deployment, and run evaluations; Codex effectively accelerated its own development. That’s both a flex and a sign of how far agentic models have come.

What it's for
GPT-5.3-Codex pulls together the best of GPT-5.2-Codex (frontier coding) and GPT-5.2 (reasoning and professional knowledge) in one model, and it’s 25% faster. So you get an agent that can take on long-running tasks — research, tool use, complex execution — and you can steer and chat with it while it works, like a colleague. No more black box: it gives frequent updates, you can ask questions or change direction mid-task, and it keeps you in the loop from start to finish.

Why it matters

Coding that leads the pack: State-of-the-art on SWE-Bench Pro (real-world software engineering across four languages) and Terminal-Bench 2.0, and strong on OSWorld (computer use) and GDPval (knowledge work across 44 occupations). It does all that with fewer tokens than before, so you can build more per session.
From code to “anything on a computer”: It’s not just write-and-review anymore. Codex can handle debugging, deployment, monitoring, PRDs, copy, user research, tests, slides, spreadsheets — the full software lifecycle and general knowledge work. In demos it built two full games (racing, diving) from scratch over millions of tokens, iterating on prompts like “fix the bug” and “improve the game.” For everyday web work, underspecified prompts now get more sensible, production-ready defaults (e.g. pricing toggles and testimonial carousels that actually look finished).
Interactive collaborator: In the Codex app, you can enable steering in Settings → General → Follow-up behavior. The model talks through what it’s doing and responds to feedback in real time.
Safety and cyber: GPT-5.3-Codex is the first model OpenAI classifies as High capability for cybersecurity under their Preparedness Framework and the first they’ve directly trained to find software vulnerabilities. They’re deploying their strongest cyber safeguards to date; some higher-risk requests may be routed to GPT-5.2. Defenders get Trusted Access for Cyber and $10M in API credits for good-faith security research.

Where you get it
GPT-5.3-Codex is available on paid ChatGPT plans wherever Codex runs: the Codex app, CLI, IDE extension, and web. API access is coming soon. The model is also 25% faster for Codex users thanks to infra and inference improvements.

Bottom line: if you want one agent that can code, reason, use tools, and do knowledge work on a computer — and you want to guide it in real time — GPT-5.3-Codex is the one to watch. It’s the first Codex that truly blurs the line between “coding assistant” and “general collaborator.”

Try it: If you’re on a paid ChatGPT plan, open the Codex app or use Codex in the IDE/CLI — GPT-5.3-Codex is there. Toggle “Follow-up behavior” in Settings → General to steer it while it works.