GPT-6 Astra
OpenAILarge Language ModelProprietaryOpenAI's most capable model, released September 2026. State-of-the-art in computer/browser use, software engineering, and cybersecurity, with a 1.05M context window. Uses opaque-recurrence reasoning whose chain of thought cannot be monitored — a point of controversy among safety researchers.
Abilities
Use Cases
Available in Tools
Availability
How to Use
Pros
- State-of-the-art on computer/browser use, terminal tasks, bug-finding, and codebase queries
- 1.05M token context window with 128K max output
- Strong science and professional multi-step workflow performance
- Cheap cached input at $1 per MTok for repeated-context agentic loops
Cons
- Opaque-recurrence reasoning — chain of thought cannot be monitored, criticized by safety researchers
- Steep long-context surcharge — requests over 272K input tokens bill at 2x/1.5x ($20/$75)
- Premium pricing at $10/$50 per MTok
- Cybersecurity capabilities gated to approved testers only
- Closed source — no model weights or training transparency
What to Use It For
Perfect For
SOTA results on computer/browser-use benchmarks — the headline capability of the release
State-of-the-art on bug-finding, terminal tasks, and large-codebase queries
Good For
Built for long-horizon professional tasks with strong tool use
Strong science performance per launch coverage and benchmarks
Not Recommended
Opaque recurrence means the reasoning process cannot be inspected — a problem for regulated domains
Try instead: Claude Opus 5, DeepSeek R1
The over-272K surcharge doubles input cost — Claude Fable 5.1 or GPT-5.6 Sol have no such cliff
Try instead: Claude Fable 5.1, GPT-5.6 Sol
Do Not Use For
At $10/1M input it is prohibitively expensive for bulk pipelines
Try instead: GPT-5.6 Sol, Claude Haiku 4.5
Closed source with no available weights — completely dependent on OpenAI API
Try instead: Llama 4 Maverick, Kimi K3