OpenAI Releases GPT-5.3-Codex, Says Internal Team Was “Blown Away” by Development Speed

OpenAI has announced GPT-5.3-Codex, a major upgrade that pushes Codex beyond writing code into handling full, end-to-end professional work on a computer. Interestingly, the launch coincides with the release of Anthropic’s Claude Opus 4.6 (rumor). You can read official details about it here.

GPT-5.3-Codex combines speed, autonomy, and real-world execution

Coming back to GPT-5.3-Codex, the new model combines the frontier coding strengths of GPT-5.2-Codex with the wider reasoning and knowledge capabilities of GPT-5.2. The result is a single agent that can research, use tools, debug, deploy, and execute long-running tasks with far less hand-holding.

OpenAI says GPT-5.3-Codex is also 25% faster, making it better suited for complex workflows that stretch across hours or even days. One standout detail is how the model helped build itself. OpenAI reveals that early versions of GPT-5.3-Codex were used internally to debug training runs, manage deployments, analyze evaluations, and even diagnose unexpected test results. In practice, the AI accelerated its own development cycle.

On benchmarks, GPT-5.3-Codex sets new highs. It leads SWE-Bench Pro, which tests real-world software engineering across multiple languages, and significantly outperforms previous models on Terminal-Bench 2.0, a measure of practical command-line and system skills. It also shows major gains on OSWorld, a benchmark focused on completing real productivity tasks inside a visual desktop environment.

The latest model expands beyond coding with a stronger security focus

Besides raw coding, OpenAI positions GPT-5.3-Codex as a general professional agent. It can build full web apps, design presentations, analyze spreadsheets, write documentation, and manage operational tasks across the software lifecycle. In demos, the model autonomously iterated on games and production-ready websites over millions of tokens.

As far as security is concerned, OpenAI classifies GPT-5.3-Codex as “high capability” for cybersecurity tasks and has deployed its most comprehensive safeguards yet, alongside new programs aimed at strengthening defensive security research.

GPT-5.3-Codex is now available to paid ChatGPT users across the Codex app, CLI, IDE extensions, and web, with API access planned next.