Model Drop

GPT-5.2-Codex

December 18, 2025 2 min read

GPT-5.2-Codex, which OpenAI shipped in Codex on December 18, 2025 (API access followed on January 14, 2026), is a version of GPT-5.2 tuned for long agentic coding jobs, with native context compaction and stronger cybersecurity skills.

Model: GPT-5.2-Codex (gpt-5.2-codex)
Maker: OpenAI
Ship date: December 18, 2025
Pricing: Paid ChatGPT plans only (Codex); no API at launch
Context: 400K tokens
Weights: Closed
Where you can use it:
- Codex on paid ChatGPT plans
- OpenAI API from January 14, 2026
When to reach for it:
- Big refactors, migrations and feature builds
- Agentic coding on native Windows
- Skip it for work that isn't coding
How it compares:
- A coding-tuned version of GPT-5.2
- Roughly level with Claude Opus 4.5 overall
What makes it different:
- Native compaction for long sessions
- Invite-only cyber access for defenders
Early reactions: Simon Willison doubted the small Terminal-Bench gain over GPT-5.2 would matter much in practice.
Read more: Introducing GPT-5.2-Codex, Addendum to GPT-5.2 System Card: GPT-5.2-Codex

The benchmarks

OpenAI didn't state a reasoning effort for its launch charts and compared GPT-5.2-Codex only with its own models (GPT-5.2, GPT-5.1, GPT-5.1-Codex-Max, GPT-5-Codex); the cyber and research rows come from the system card (PaperBench at Extra High effort, CTFs scored pass@12 over 16 rollouts, CVE-Bench pass@1 over 3 rollouts on 34 of 40 challenges).

OpenAI's numbers

SWE-bench Pro (harder multi-language fixes): 56.4%. +0.8 over GPT-5.2 (55.6%), +5.6 over GPT-5.1 (50.8%).
Terminal-Bench 2.0 (command-line agent tasks): 64.0%. +1.8 over GPT-5.2 (62.2%), +5.9 over GPT-5.1-Codex-Max (58.1%).
Professional Capture-the-Flag Challenges (pro hacking puzzles): 88.1%. +5.7 over GPT-5.2 (82.4%), +12.5 over GPT-5.1-Codex-Max (75.6%), +37.6 over GPT-5-Codex (50.5%).
CVE-Bench (exploiting real web bugs): 87%. +7 over GPT-5.1-Codex-Max (80%), +18 over GPT-5.2 (69%).
Cyber Range (end-to-end network attacks, combined pass rate): 72.7%. +9.1 over GPT-5.2 (63.6%), −9.1 under GPT-5.1-Codex-Max (81.8%).
PaperBench-10 (replicating AI papers): 43%. +3 over GPT-5.1-Codex-Max (40%), +4 over GPT-5.2 (39%).
MLE-Bench-30 (Kaggle ML contests): 10%. −7 under GPT-5.1-Codex-Max (17%), −6 under GPT-5.2 (16%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 28.5. +0.5 over Gemini 3 Pro (28.0), −0.6 under Claude Opus 4.5 (29.1), −1.9 under GPT-5.2 (30.4).