Introducing GPT-5.2-Codex (OpenAI) Dec 18, 2025
GPT-5.2 Codex
GPT-5.2 Codex places in the top three on 3 benchmarks.
41
Benchmarks with a score
0
Ranked first, out of five or more models
3
In the top three
OpenAI Dec 2025
GPT-5.2 Codex
Closed weights
149.9
#95 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- SWE-Bench Pro 41% #11 /24
- Terminal-Bench 2.0 66.5% #7 /43
- PostTrainBench 17.2% #14 /23
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Wer wird Millionär? LLM Benchmark | 200000.0 | #12 of 77 | Official leaderboard | |
| BullshitBench | 36.4% | #42 of 107 | Official leaderboard | |
| BullshitBench V2 | 45% | #60 of 129 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| OpenAI PRs | 55% | #2 of 12 | Lab model card | |
| SWE-bench Verified | 72.8% | #11 of 60 | Official leaderboard | |
| OpenAI-Proof Q&A | 8.00% | #2 of 9 | Lab model card | |
| OpenHands Index | 58.3% | #11 of 33 | Official leaderboard | |
| ALE-Bench | 1,686 | #47 of 126 | Official leaderboard | |
| PaperBench | 43% | #6 of 15 | Lab model card | |
| OTelBench | 15.9% | #6 of 14 | Official leaderboard | |
| SWE-Bench Pro | 41% | #11 of 24 | Official leaderboard | |
| SlopCodeBench | 21.9% | #11 of 22 | Official leaderboard | |
| SWE-Bench Pro Private | 27.7% | #6 of 12 | Official leaderboard | |
| Vibe Code Bench | 37.9% | #73 of 120 | Official leaderboard | |
| PostTrainBench | 17.2% | #14 of 23 | Official leaderboard | |
| MLE-bench 30 | 10% | #6 of 8 | Lab model card | |
| SWE-bench Multilingual | 66.3% | #10 of 13 | Official leaderboard | |
| WebDev Arena | 1,339 | #84 of 105 | arena.ai, via Epoch AI | |
| Code Arena | 1,338 | #105 of 128 | Official leaderboard | |
| APEX-SWE | 30.8% | #54 of 65 | Official leaderboard | |
| DesignArena Agentic Overall | 1,032 | #65 of 74 | Official leaderboard | |
| DesignArena React Native | 1,084 | #50 of 56 | Official leaderboard | |
| DesignArena Agentic Image to HTML | 1,139 | #10 of 11 | Official leaderboard | |
| DesignArena Frontend Web App | 1,032 | #51 of 53 | Official leaderboard | |
| DesignArena Full Stack | 972 | #48 of 49 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SnakeBench | 32 | #24 of 188 | Official leaderboard | |
| Terminal-Bench 2.0 | 66.5% | #7 of 43 | Official leaderboard | |
| WeaveBench | 6.10% | #6 of 9 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| EnterpriseBench: CoreCraft | 20.1% | #44 of 51 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MineBench | 1,052 | #65 of 75 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BinaryAudit | 45.6% | #3 of 20 | Official leaderboard | |
| Capture the Flag (Professional) | 88.1% | #4 of 21 | Lab model card | |
| ProtocolQA Open-Ended | 39% | #10 of 27 | Lab model card | |
| CVE-Bench | 87% | #6 of 14 | Lab model card | |
| Tacit Knowledge and Troubleshooting | 73% | #13 of 30 | Lab model card | |
| StrongREJECT | 93.3% | #10 of 19 | Lab model card | |
| TroubleshootingBench | 34% | #12 of 20 | Lab model card | |
| Cyber Range | 72.7% | #6 of 10 | Lab model card | |
| Destructive action avoidance | 76% | #6 of 8 | Lab model card | |
| Multimodal Troubleshooting Virology | 48% | #22 of 27 | Lab model card | |
| CyScenarioBench | 0% | #14 of 17 | Lab model card |
Its lab's numbers put it at 149.9, number 95 of 404
The Self-Reported Index fits one score to the 6 capability results OpenAI reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 146.4 to 153.4. How we compute it.
- Terminal-Bench 2.0 64
- SWE-Bench Pro 56.4
- OpenAI PRs 55
- PaperBench 43
- MLE-bench 30 10
- OpenAI-Proof Q&A 8
Its lab's numbers next to 2 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| SWE-Bench Pro | 56.4% SWE-Bench Pro | 41% Leaderboard, SWE-agent | +15 pts |
| Terminal-Bench 2.0 | 64% Terminal-Bench 2.0 | 66.5% Leaderboard, Deep Agents | −2.5 pts |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 16 of them aren't in the launch post.
Addendum to GPT-5.2 System Card: GPT-5.2-Codex Dec 18, 2025
Safety
Robustness
- StrongREJECT 93.3
Alignment
Bio and chem risk
Cyber risk
AI R&D risk
- OpenAI PRs 55
- MLE-bench 30 10
- PaperBench 43
- OpenAI-Proof Q&A 8