OpenAI has introduced GPT-5.3-Codex, its most capable agentic coding model, which integrates the frontier coding performance of GPT-5.2-Codex with the professional reasoning and knowledge capabilities of GPT-5.2. This release marks a significant acceleration in agentic capabilities, supporting long-running tasks involving research, tool usage, and complex execution. The model achieved new state-of-the-art results on key industry benchmarks, including SWE-Bench Pro (56.8% accuracy) and Terminal-Bench 2.0 (77.3% accuracy), and demonstrated superior performance on OSWorld. Internally, the model was instrumental in its own development, deployment, and debugging processes. It also exhibits enhanced professional knowledge work capabilities, matching GPT-5.2 on GDPval. Given its advanced capabilities, GPT-5.3-Codex is classified as "High capability" for cybersecurity, leading to the deployment of comprehensive safety mitigations and the launch of a Trusted Access program focused on accelerating cyber defense research. The model offers a 25% increase in speed for Codex users.
GPT-5.3-Codex Strategic Capability Summary
Model Synthesis and Speed: GPT-5.3-Codex combines the high-fidelity coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2. It operates 25% faster for Codex users due to infrastructure and inference stack improvements.
Agentic Task Execution: The model is designed to handle long-running, complex tasks that require research, external tool use, and intricate execution steps. It supports interactive collaboration, providing frequent updates and allowing users to steer its progress in real-time without losing context.
Frontier Coding Performance:
Achieves State-of-the-Art (SOTA) on SWE-Bench Pro (56.8% accuracy), an industry-relevant evaluation spanning four programming languages.
Achieves SOTA on Terminal-Bench 2.0 (77.3% accuracy), measuring essential terminal skills.
Enhanced Web Development: The model demonstrates the ability to autonomously build highly functional, complex applications and games from scratch over multi-day iteration cycles, requiring millions of tokens (demonstrated via a racing game and a diving game). It defaults to more functional and production-ready outputs for underspecified prompts (e.g., pricing tables, testimonial carousels).
Broader Professional Knowledge Work: The agent’s capabilities extend beyond code generation to support the full software lifecycle (debugging, deployment, monitoring) and general professional tasks such as creating slide decks, analyzing data in spreadsheets, and writing PRDs. It matches the performance of GPT-5.2 on the GDPval benchmark for knowledge-work tasks.
Superior Computer Use: Performance on the OSWorld-Verified benchmark, which measures productivity tasks in a visual desktop environment, dramatically improved to 64.7% accuracy (up from 38.2% for GPT-5.2-Codex).
Internal Development Acceleration: Early versions of GPT-5.3-Codex were used by the Codex team to accelerate its own development, specifically in debugging training runs, managing deployment, analyzing interaction quality, and building rich analytical applications for researchers.
Cybersecurity Classification and Mitigation:
The model is the first to be classified as High capability for cybersecurity tasks under OpenAI’s Preparedness Framework.
It was directly trained to identify software vulnerabilities.
A comprehensive safety stack, including automated monitoring, safety training, and trusted access protocols, has been deployed to mitigate dual-use risks.
OpenAI is launching Trusted Access for Cyber, a pilot program to accelerate cyber defense research, supported by a $10M commitment in API credits for open source and critical infrastructure security research.
Availability: GPT-5.3-Codex is available immediately through paid ChatGPT plans via the Codex app, CLI, IDE extension, and web interface. API access is expected to be safely enabled soon.