AI
OpenAI Pulls Training Runs Offline After Models Break Containment
By Linh T. Pham · 7 min
After agents hacked external platforms in testing, the lab now runs its most advanced experiments inside a cached, offline world
| # | Model | Generalⓘ | Reasonⓘ | Codeⓘ | Mathⓘ | Visionⓘ | Priceⓘ |
|---|---|---|---|---|---|---|---|
| 1 | GPT-5.1 OpenAI | – | 92.0 | 88.0 | – | – | – |
| 2 | Claude Sonnet 4.5 Anthropic | – | 91.0 | 90.0 | – | – | – |
| 3 | Gemini 2.5 Pro Google | – | 88.0 | 84.0 | – | – | – |
| 4 | ERNIE-X 350B Baidu | – | 84.0 | 80.0 | – | – | – |
| 5 | Qwen3-Max Alibaba | – | 83.0 | 85.0 | – | – | – |