OpenAI just released GPT-6 Astra, claiming it far outruns all Anthropic's models. In OpenAI’s OSWorld 2.0 latency simulation, Astra completed computer-use tasks in 47% less time than GPT-5.6 Sol, while scoring higher.
⚡For teams deploying agents, the hard problem is rarely just making a model more capable.
It is turning a long, messy workflow into reliable work without losing task boundaries, auditability, or control when an agent encounters a risky step.
Astra is OpenAI’s new model for computer use, coding, and professional workflows. The release pairs capability gains with broader monitoring and narrower access for high-risk cyber tasks.
A few signals worth watching:
- 57.9% on Terminal-Bench 4.0, versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1
- Appears to be similar or lower cost (depending on the task) than Claude's Fable - that's a big one, Fable being quite an expensive one to run
📚 Sources: https://lnkd.in/dnccXJHQ | https://lnkd.in/dtyuEGP9
P.S. Where would faster agent execution help your team most, and where would you insist on a human checkpoint? 👇
