A bit underneath two months after releasing its final main mannequin household to the general public, GPT-5.6 Sol, Terra and Luna, OpenAI is again with a wholly new one: GPT-6 Astra, what it claims is “essentially the most clever and aligned mannequin on the planet.” Following the corporate’s resolution to sluggish frontier mannequin improvement in August after one in all its fashions hacked AI platform Hugging Face, it could even be the final main AI mannequin it releases for some time.
GPT-6 Astra excels at computer-use and looking, and dealing with duties within the domains of software program engineering, cybersecurity, science and basic skilled work, in keeping with OpenAI. A flashy demo video the corporate launched alongside the launch reveals Astra dealing with every part from 3D modeling to constructing slideshows, and infrequently taking good care of a number of duties throughout totally different domains on the identical time (like ordering meals whereas coding a recreation). Key to Astra’s attraction is its means to deal with these multi-step workflows in your laptop and out and in of your browser. It is also allegedly in a position to do these advanced duties with “sturdy visible judgment,” OpenAI says, and with out drifting from its unique instructions or immediate.
As with different mannequin launches, the corporate has a brand new tranche of benchmarks to justify Astra’s improved talents, together with a 98.6% rating on ARC-AGI-3, one of many trade’s benchmarks for measuring an AI’s means to unravel unfamiliar issues. VentureBeat notes that the rating has some caveats, specifically that AIs which were put by the benchmark aren’t all configured in the identical approach, and totally different system structure (like whether or not they have persistent reminiscence, for instance) can produce totally different outcomes. Different enhancements are extra clear-cut: Astra scored 57.7% on Terminal Bench 4.0, a coding benchmark, and 59.3% on the Agent’s Final Examination, a benchmark for measuring an AI’s agentic capabilities, notably greater than GPT-5.6 Sol. OpenAI included much more benchmarks in its announcement, however briefly, Astra is now on the prime of most leaderboards by way of efficiency.
That raises the bigger query of whether or not the AI mannequin can be utilized and deployed responsibly. For instance, Astra is especially expert at cybersecurity. OpenAI says the mannequin scored an ideal rating on ExploitBench, a benchmark for figuring out a mannequin’s means to use software program vulnerabilities, which was an enormous soar up from GPT-5.6 Sol’s rating of 78.5%. The corporate additionally examined Astra on SRE-Bench, a benchmark that measures a mannequin’s means to reverse engineer software program binaries with out uncooked supply code, and located that “Astra solved 88.0% of duties in a single try and 99.2% inside 4 makes an attempt, in contrast with 55.9% and 68.7% for GPT-5.6 Sol, respectively.”
Astra’s improved alignment — masking every part from its means to observe templates to its means to talk extra transparently — ought to enhance security, however OpenAI says it is also constructed the mannequin to not adjust to superior cybersecurity duties. The corporate says stronger protections will even assist Astra higher stand up to jailbreak requests whereas additionally enhancing the corporate’s means to observe misuse.
OpenAI says that GPT-6 Astra is “rolling out at this time to a restricted set of organizations and over the approaching days will grow to be out there to all ChatGPT Plus, Professional, Enterprise, and Enterprise customers, in addition to by the OpenAI API and AWS.” For builders utilizing the OpenAI API, Astra prices $10 per million enter tokens and $50 per million output tokens.
That places Astra on the costly finish of AI fashions, although OpenAI argues tokens won’t be the perfect metric for companies seeking to purchase entry. “What you truly need, and I feel the market is beginning to actually get up to, is the value per activity,” OpenAI President Greg Brockman advised VentureBeat. “It is nearly: are you able to get the factor carried out for an applicable value at applicable pace?” OpenAI’s huge promise is that for all kinds of companies, Astra can.
