Astra
OpenAI's next major model with critical cybersecurity capabilities
Context window
1050000
Input / 1M tokens
10
Output / 1M tokens
50
Provider
OpenAI
Data verified 2026-09-05
Astra is OpenAI's next major model family announced August 1, 2026. It is trained to enable persistent agents and demonstrated breakthrough performance on mathematical and theoretical computer science problems. The model has reached OpenAI's first "Critical" cybersecurity capability threshold, able to identify and develop working zero-day exploits without human intervention.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Limited preview with alpha testers and Daybreak Blue program for defensive cybersecurity purposes. Not yet publicly available.
Strengths
- ✓Advanced mathematical and theoretical problem-solving
- ✓Critical cybersecurity capabilities (zero-day discovery and exploit development)
- ✓Multi-agent long-horizon reasoning and planning
- ✓Test-time compute for complex scientific reasoning
Best for developers who...
When to choose it (and when not to)
Reach for Astra when...
- →When you need cutting-edge mathematical problem-solving
- →For authorized defensive cybersecurity work through Daybreak Blue
Look elsewhere if...
- ✕For general-purpose tasks (use GPT-5.6 Sol instead)
- ✕Before official public release and authorization
How to use it
Quickstart
Python# Astra not yet publicly available for API useModel is in limited preview testing phase only
API model id: gpt-6-astra
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| ExploitBench | 100% | Internal evaluation; Astra scored 100% on cybersecurity exploitation benchmark |
| Mathematics (open problems) | 10 problems solved | Demonstrated solutions to 10 long-standing open problems in mathematics and theoretical computer science |
| Cyber refusal rate | 91.5% | Refuses 91.5% of cyber-jailbreak requests vs 59% for GPT-5.6 Sol |
| Terminal-Bench 4.0 | 57.9 | vs Sol 37.3% |
| DeepSWE v1.1 | 74.1 | vs Sol 72.7% |
| OSWorld 2.0 | 72.6 | vs Sol 65.7% |
| FrontierMath Tier 4 | 97.6 | vs Sol 83.0% |
Compare Astra
Compare Astra with any other model
Build a comparison →All model comparisons →