GPT-5-Codex

// Overview

GPT-5-Codex is the first Codex-branded model in OpenAI's GPT-5 generation, announced in September 2025. It is a version of GPT-5 further optimized for agentic coding inside Codex, trained on complex, real-world software engineering work rather than general chat: building full projects from scratch, adding features and tests, debugging, performing large-scale refactors, and conducting code reviews. OpenAI recommends GPT-5-Codex specifically for agentic coding tasks in Codex or Codex-like environments, while keeping GPT-5 as the general-purpose model.

GPT-5-Codex carries a 400K-token context window and can generate up to 128K output tokens, and it accepts both text and image input so it can read screenshots, diagrams, and UI surfaces shared during a coding session. A defining trait is its ability to dynamically scale how long it thinks: it answers quickly on small, interactive requests but can keep working independently for long stretches on large tasks. During OpenAI's testing, GPT-5-Codex worked on its own for more than seven hours at a time on large, complex tasks, iterating on its implementation, fixing test failures, and ultimately delivering a working result.

The model became available to developers on September 23, 2025, through the OpenAI Responses API (model ID gpt-5-codex), in addition to ChatGPT subscriptions. It runs everywhere Codex does — the Codex CLI, IDE extensions, the web, GitHub code review, and the ChatGPT iOS app — and is the default for cloud tasks and code review. On SWE-bench Verified it scores 74.5%, and on OpenAI's internal code-refactoring benchmark it reaches 51.3%, a large jump over the base GPT-5 model's 33.9%.

// Benchmarks

1. SWE-bench Verified74.5%
2. Code refactoring (OpenAI internal benchmark)51.3%

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

// Pricing

Pricing source ↗

// Strengths

• Long-horizon agentic coding — worked independently for 7+ hours on large tasks in OpenAI's testing
• Dynamically scales reasoning time: fast on interactive requests, persistent on complex ones
• Strong on large-scale code changes such as refactors, validating with dependency analysis and tests
• Image input — reads screenshots, diagrams, and UI surfaces during a coding session
• 400K-token context window for reasoning over large codebases
• Runs everywhere Codex does: CLI, IDE extension, web, GitHub code review, and the ChatGPT iOS app

// Best for

• Reach for it for long, independent execution of complex engineering tasks inside Codex (CLI, IDE, or cloud).
• Reach for it for large-scale refactors and migrations that need planning, dependency analysis, and test runs.
• Reach for it for automated code review on GitHub pull requests.
• Reach for it when a coding agent needs to interpret a UI screenshot or diagram as part of the loop.

// 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.