Model Drop

GPT-5.3-Codex

February 5, 2026 1 min read

GPT-5.3-Codex, which OpenAI shipped on February 5, 2026, is a Codex agent model that combines GPT-5.2-Codex's coding with GPT-5.2's reasoning and professional knowledge in one faster model.

Model: GPT-5.3-Codex (gpt-5.3-codex)
Maker: OpenAI
Ship date: February 5, 2026
Pricing: Paid ChatGPT plans only (Codex); no API at launch
Context: 400K tokens
Where you can use it:
- Codex app, CLI, IDE extension and web
- OpenAI API from February 24
When to reach for it:
- Long, terminal-heavy coding jobs
- Desktop tasks run from Codex
- Skip it for creative greenfield builds
How it compares:
- Launched the same day as Claude Opus 4.6
- Roughly level with Opus 4.6 overall
What makes it different:
- Helped debug its own training run
- OpenAI's first "High" cyber-risk model
Early reactions: Developers rated it strong at code review but weaker on creative, greenfield work, and many paired it with Claude Opus 4.6.
Read more: Introducing GPT-5.3-Codex, GPT-5.3-Codex system card

The benchmarks

OpenAI ran every model at xhigh effort (GPT-5.2 at high for GDPval) and compared GPT-5.3-Codex only with its own GPT-5.2-Codex and GPT-5.2; it trails neither on any row it published, and GDPval is a tie.

OpenAI's numbers

SWE-bench Pro (harder multi-language fixes): 56.8%. +0.4 over GPT-5.2-Codex (56.4%), +1.2 over GPT-5.2 (55.6%).
Terminal-Bench 2.0 (command-line agent tasks): 77.3%. +13.3 over GPT-5.2-Codex (64.0%), +15.1 over GPT-5.2 (62.2%).
OSWorld-Verified (desktop computer use): 64.7%. +26.5 over GPT-5.2-Codex (38.2%), +26.8 over GPT-5.2 (37.9%).
GDPval (expert knowledge work, wins or ties): 70.9%. +0.0 over GPT-5.2 (70.9%).
Cybersecurity Capture The Flag Challenges (hacking puzzles): 77.6%. +10.2 over GPT-5.2-Codex (67.4%), +9.9 over GPT-5.2 (67.7%).
SWE-Lancer IC Diamond (freelance coding jobs): 81.4%. +5.4 over GPT-5.2-Codex (76.0%), +6.8 over GPT-5.2 (74.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 32.5. +0.6 over Claude Opus 4.6 (31.9), +4.0 over GPT-5.2-Codex (28.5), +4.5 over Gemini 3 Pro (28.0).