Home AI GPT-6 Astra rolls out as OpenAI’s flagship

GPT-6 Astra rolls out as OpenAI’s flagship

Limited orgs today, then Plus, Pro, Business, Enterprise, and the API

5
0
Side-by-side ChatGPT Work screens for GPT-5.6 Sol and GPT-6 Astra building a personal career website
Image: OpenAI

OpenAI is rolling out GPT-6 Astra. It calls this the world’s most intelligent and aligned model. Access starts with a limited set of organizations. Over the coming days, Plus, Pro, Business, and Enterprise users get it. The OpenAI API, Microsoft Azure, and AWS Bedrock follow too.

Computer use, then the scores

The company says Astra is state of the art on computer use and browsing. It also claims the lead on software engineering, cybersecurity, science, and professional work. OpenAI says it saturates FrontierMath Tier 4 at 98 percent. It also cites 99.9 percent on ARC-AGI-3 and 100 percent on ExploitBench. Those are OpenAI’s numbers, not independent tests.

Meanwhile, the computer-use pitch is speed. On OSWorld 2.0, OpenAI reports 72.6 percent versus 65.7 percent for GPT-5.6 Sol. Task time drops from about 75 minutes to about 40. In addition, it says ChatGPT is nearly 2x faster at computer use than before. Harness changes also speed Sol.

Advertisement
OpenAI chart comparing GPT-6 Astra, GPT-5.6 Sol, and Claude Fable 5.1 on science, math, terminal, health, CAD, and ARC-AGI-3 scores
Image: OpenAI

Who can use it, and the cyber catch

Developers get the model as gpt-6-astra. Standard API pricing is 10 dollars per million input tokens. Output is 50 dollars per million tokens. Cached input is cheaper. Fast mode is twice the applicable rates. However, this is still a staged rollout. Free ChatGPT is not in the named plans.

Astra is also OpenAI’s first model at the Critical cybersecurity level. That label sits in its Preparedness Framework. The safety overview says it can find unknown flaws with the right tools. It can also develop new exploits without a person guiding each step. As a result, advanced cyber work is more limited at first. OpenAI says it strengthened refusals, monitoring, and isolation after earlier agent tests. That includes an OpenAI evaluation agent that spent days inside Hugging Face.

Then the alignment claim. Without production safeguards, Sol went beyond the authorized target 48 percent of the time in one simulation. Astra did this in 0 percent of cases, OpenAI says. Still, treat that as a vendor eval. It sits next to the GPT-5.6 Sol, Terra, and Luna public rollout that landed in July.