GPT-5.3-Codex

// Overview

GPT-5.3-Codex is OpenAI's coding-specialized model, released February 5, 2026, as part of the Codex line. OpenAI describes it as its most capable agentic coding model to date: it merges the frontier coding performance of GPT-5.2-Codex with the broader reasoning and professional-knowledge skills of GPT-5.2 into a single model that runs about 25% faster. It is tuned for long-horizon software-engineering work — research, tool use, terminal commands, repository search, and multi-step debugging — and lets you steer and interact with it mid-task without losing context.

GPT-5.3-Codex accepts text and image input and returns text, with a 400K-token context window and up to 128K output tokens. In the API it is exposed as the model ID gpt-5.3-codex and supports low, medium, high, and xhigh reasoning-effort settings along with streaming, function calling, and structured outputs. Its knowledge cutoff is August 31, 2025.

OpenAI notes that GPT-5.3-Codex was the first of its models to be instrumental in creating itself — the Codex team used early versions to help debug its own training run, support deployment, and diagnose evaluation results. It was also the first OpenAI model classified as 'High capability' for cybersecurity under the company's Preparedness Framework. The model has since been superseded by later Codex/GPT releases (GPT-5.4 and GPT-5.5) and is deprecated as a user-selectable Codex model when signed in with ChatGPT, though it remains available via the API.

// Benchmarks

1. SWE-Bench Pro56.8%
2. Terminal-Bench 2.077.3%
3. OSWorld-Verified64.7%
4. GDPval (win-or-tie rate)70.9%
5. Cybersecurity Capture The Flag77.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Pricing source ↗

// Strengths

• Sets an industry-high score on SWE-Bench Pro, a contamination-resistant real-world software-engineering benchmark spanning four languages
• Strong agentic and computer-use performance: leading Terminal-Bench 2.0 and OSWorld-Verified results
• Long-horizon autonomy — research, tool use, and complex multi-step execution with interactive mid-task steering
• Large 400K-token context with 128K-token output for working across big codebases and long sessions
• About 25% faster than GPT-5.2-Codex while also folding in GPT-5.2's general reasoning and professional knowledge
• Tunable reasoning effort (low / medium / high / xhigh) to trade latency for depth

// Best for

• Autonomous, long-running software-engineering tasks inside the Codex app, CLI, IDE extension, and web
• Repository-wide code search, refactoring, and multi-file changes
• Running and reasoning over terminal commands and build/test output
• Multi-step debugging, diagnosing test failures, and triaging evaluation results
• Agentic computer-use and tool-driven workflows
• Knowledge-work tasks such as document and spreadsheet generation (GDPval-style work)

// How to access

// Codex — every version

The full lineage of the Codex line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.