Seven weeks. That’s roughly how long the gap was between OpenAI shipping GPT-5.6 Sol, Terra and Luna to the public and the arrival of an entirely new model family, GPT-6 Astra, which the company describes as “the most intelligent and aligned model in the world.”
It might also be the last big one for a while. OpenAI decided in August to slow frontier model development after one of its models hacked the AI platform Hugging Face, and Astra lands in the shadow of that decision.
What Astra is actually built to do
The pitch here isn’t better prose. It’s your computer. OpenAI says Astra excels at computer-use and browsing, plus work in software engineering, cybersecurity, science and general professional tasks.
The company’s launch demo shows the model doing 3D modeling and building slideshows, frequently juggling several tasks across different domains at once, including ordering food while coding a game.
The load-bearing claim is that Astra can hold a multi-step workflow together on your machine, moving in and out of the browser without losing the plot. OpenAI says it does this with “strong visual judgment” and without drifting from the original prompt or directions. Allegedly.
The benchmark that needs an asterisk
OpenAI brought numbers, as it always does. The headline figure is 98.6% on ARC-AGI-3, an industry benchmark for solving unfamiliar problems.
Treat that one carefully. Models put through the benchmark aren’t all configured identically, and differences in system architecture, such as whether a model has persistent memory, can move the result.
The rest is cleaner. Astra scored 57.7% on the coding benchmark Terminal Bench 4.0 and 59.3% on the Agent’s Last Exam, which measures agentic ability, both notably higher than GPT-5.6 Sol. OpenAI listed more benchmarks in its announcement, and the short version is that Astra now sits at or near the top of most leaderboards.
The cybersecurity numbers are the story
Here’s where the launch gets genuinely uncomfortable. Astra scored a perfect score on ExploitBench, which measures a model’s ability to exploit software vulnerabilities. GPT-5.6 Sol scored 78.5% on the same test.
On SRE-Bench, which measures reverse engineering of software binaries without raw source code, OpenAI reports that “Astra solved 88.0% of tasks in a single attempt and 99.2% within four attempts, compared with 55.9% and 68.7% for GPT-5.6 Sol, respectively.”
Read those two together and you get a model that went from competent to near-perfect at finding and exploiting software flaws in under two months of release cadence. That’s the responsible-deployment question sitting in plain view.
OpenAI’s answer is alignment work, covering everything from following templates to speaking more transparently, and a model built to refuse advanced cybersecurity tasks outright. The company says stronger protections should also help Astra hold up against jailbreak attempts while making misuse easier to monitor. Refusal training and a perfect ExploitBench score are living in the same model, and how well that holds is something users will find out before OpenAI publishes anything about it.
Who gets it, and what it costs
Astra is “rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS,” OpenAI said.
API pricing is $10 per million input tokens and $50 per million output tokens. That’s the expensive end of the market, and OpenAI knows it, which is why the company would rather you stopped counting tokens.
“What you actually want, and I think the market is starting to really wake up to, is the price per task,” OpenAI President Greg Brockman said. “It’s just about: can you get the thing done for an appropriate cost at appropriate speed?”
That framing only works if Astra finishes the job on the first pass. At $50 per million output tokens, a model that needs three attempts at your workflow is a model you can’t afford, and the SRE-Bench split between one attempt and four is exactly the gap that decides your bill. Run your own agentic task through it before you move a team over, and count retries, not tokens.
