This Is the First AI That Helped Build Itself - Meet GPT-5.3-Codex

GPT-5.3-Codex, an advanced model, enhances coding performance and reasoning, operating 25% faster than its predecessor. It excels in industry benchmarks, supports the software lifecycle, and can autonomously build complex applications. The model is available on multiple platforms with plans for API access.

OpenAI has released GPT-5.3-Codex, an updated version of GPT-5.2-Codex (itself a specialized variant of GPT-5.2 optimized for agentic coding). The new model runs 25% faster and combines stronger frontier-level coding performance with improved reasoning and professional knowledge. It is designed for long-running tasks that mix research, tool use, and complex execution, and it delivers strong results on industry benchmarks such as SWE-Bench Pro and Terminal-Bench.

Beyond raw coding ability, GPT-5.3-Codex represents a shift in scope. It is positioned not just as a code-writing assistant, but as a general-purpose, computer-using agent capable of operating across terminals, browsers, IDEs, and desktop environments. This broader capability is reflected in its performance on agentic benchmarks like OSWorld, which measure real-world computer use rather than isolated code generation. The model can autonomously build and iterate on complex applications, including full web games, and it shows improved intent understanding, producing more complete and production-ready results from underspecified prompts.

GPT-5.3-Codex is built to support the full software lifecycle. In addition to writing and reviewing code, it assists with debugging, deploying, monitoring, writing product requirement documents (PRDs), creating tests, analyzing metrics, and performing user research. Its professional knowledge capabilities are validated by strong results on GDPval, an evaluation that measures performance across well-specified knowledge-work tasks spanning dozens of occupations.

A key change in this release is how developers interact with the model. GPT-5.3-Codex supports interactive steering, allowing users to guide the agent while it is working. Instead of waiting for a final output, users can receive progress updates, ask questions, adjust direction, and provide feedback without losing context. This positions Codex more like a collaborative teammate than a fire-and-forget automation tool.

Notably, GPT-5.3-Codex was also used extensively in its own development. Early versions helped debug training runs, analyze evaluation anomalies, build data pipelines, and optimize deployment infrastructure. During launch, the model assisted with scaling GPU clusters and stabilizing latency under traffic surges. This process marks a rare example of an AI system materially accelerates its own research and deployment processes.

Cybersecurity is another major focus of this release. GPT-5.3-Codex is the first OpenAI model classified as “High capability” for cybersecurity-related tasks under the Preparedness Framework. While OpenAI reports no evidence that it can fully automate end-to-end cyber attacks, the model has been directly trained to identify software vulnerabilities. As a result, it is deployed with strengthened safeguards, including safety training, automated monitoring, trusted access controls, and enforcement pipelines. OpenAI has also committed $10 million in API credits to accelerate defensive cybersecurity research and ecosystem resilience.

GPT-5.3-Codex is available today through paid ChatGPT plans, everywhere Codex is supported: the app, CLI, IDE extension, and web. API access is planned but not yet generally available. The model is trained and served on NVIDIA systems.

• https://openai.com/index/introducing-gpt-5-3-codex...