OpenAI GPT-6 Astra Launch - Internal Computer Use Safety Benchmark: leaderboard
Alignment table of OpenAI's GPT-6 Astra launch post (2026-09-03): rate of unintended consequences on computer-use tasks adversarially selected to elicit misbehavior. Percent, lower is better. Footnote 16: third-party models run in a simpler research setup than Codex, and the no-confirmation scenario is an internal research configuration.
Metric: Unintended-consequence rate (%). Source: openai.com. Source: openai.com. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 2.4 | #6 |
| 2 | GPT-6.1 Sol (xHigh) | 4.32 | #7 (GPT-6.1 Sol) |
| 3 | Claude Fable 5.1 | 9.5 | #10 |
| 4 | Claude Opus 5 | 11.5 | #16 |
| 5 | GPT-6 Luna (xHigh) | 13.7 | #61 (GPT-6 Luna) |
| 6 | GPT-6 Sol (xHigh) | 17.39 | #19 (GPT-6 Sol) |
| 7 | Claude Fable 5 | 18.3 | #20 |
| 8 | GPT-5.6 Sol | 22 | #22 |
No result here: #3 Claude Opus 5.5, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=openai-gpt-6-astra-launch-internal-computer-use-safety-benchmark · How It Works · Data refreshed daily, snapshot 2026-11-05.