The new frontier model is particularly competent at agentic, computer-use tasks.
A little under two months after releasing its last major model family to the public, GPT-5.6 Sol, Terra and Luna, OpenAI is back with an entirely new one: GPT-6 Astra, what it claims is “the most intelligent and aligned model in the world.” Following the company’s decision to slow frontier model development in August after one of its models hacked AI platform Hugging Face, it may also be the last major AI model it releases for a while.
GPT-6 Astra excels at computer-use and browsing, and handling tasks in the domains of software engineering, cybersecurity, science and general professional work, according to OpenAI. A flashy demo video the company released alongside the launch shows Astra handling everything from 3D modeling to building slideshows, and often taking care of multiple tasks across different domains at the same time (like ordering food while coding a game). Key to Astra’s appeal is its ability to handle these multi-step workflows on your computer and in and out of your browser. It’s also allegedly able to do those complex tasks with “strong visual judgment,” OpenAI says, and without drifting from its original directions or prompt.
As with other model launches, the company has a new tranche of benchmarks to justify Astra’s improved abilities, including a 98.6% score on ARC-AGI-3, one of the industry’s benchmarks for measuring an AI’s ability to solve unfamiliar problems. VentureBeat notes that the score has some caveats, namely that AIs that have been put through the benchmark aren’t all configured in the same way, and different system architecture (like whether they have persistent memory, for example) can produce different results. Other improvements are more clear-cut: Astra scored 57.7% on Terminal Bench 4.0, a coding benchmark, and 59.3% on the Agent’s Last Exam, a benchmark for measuring an AI’s agentic capabilities, notably higher than GPT-5.6 Sol. OpenAI included even more benchmarks in its announcement, but in short, Astra is now at the top of most leaderboards in terms of performance.
That raises the larger question of whether the AI model can be used and deployed responsibly. For example, Astra is particularly skilled at cybersecurity. OpenAI says the model scored a perfect score on ExploitBench, a benchmark for determining a model’s ability to exploit software vulnerabilities, which was a big jump up from GPT-5.6 Sol’s score of 78.5%. The company also tested Astra on SRE-Bench, a benchmark that measures a model’s ability to reverse engineer software binaries without raw source code, and found 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.”
Astra’s improved alignment — covering everything from its ability to follow templates to its ability to speak more transparently — should improve safety, but OpenAI says it’s also built the model to not comply with advanced cybersecurity tasks. The company says stronger protections will also help Astra better withstand jailbreak requests while also improving the company’s ability to monitor misuse.
OpenAI says that GPT-6 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.” For developers using the OpenAI API, Astra costs $10 per million input tokens and $50 per million output tokens.
That puts Astra on the expensive end of AI models, though OpenAI argues tokens might not be the best metric for businesses looking to buy access. “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 told VentureBeat. “It’s just about: can you get the thing done for an appropriate cost at appropriate speed?” OpenAI’s big promise is that for a wide variety of businesses, Astra can.
