OpenAI throws Astra into the top-tier model ring

After a brief training pause to reflect on the inadequacy of its AI security measures, OpenAI has released GPT-6 Astra, initially for those in its Trusted Access Program. Assuming all goes well and there are no more incidents like the Hugging Face hack, the AI biz in a few days will allow Plus, Pro, Business, and Enterprise subscribers to try out what the company describes as "our most intelligent model yet, with state-of-the-art performance in computer use, browsing, software engineering, science, and professional work." GPT-6 Astra arrives just two days after rival Anthropic debuted Fable 5.1, at the same eye-watering price: $10 per million input tokens and $50 per million output tokens. Astra, OpenAI says, is the first of its models to reach the "Critical" level for cybersecurity capabilities under its Preparedness Framework. The 22-page policy document has been panned [PDF] by researchers, who argue that it "allows OpenAI's CEO to deploy even more dangerous capabilities, especially if other AI developers do so." And here we are. OpenAI's cautionary but upbeat assessment of Astra describes it as "a significant step up in cyber capabilities." "This means that, with the right tools and access, GPT6 Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step," the company said. OpenAI goes on to argue not to worry since Astra is more robust than prior models, better aligned, more comprehensively monitored, and safer in high-risk scenarios. "[W]e built a new evaluation informed by the Hugging Face incident that evaluates whether a model facing a difficult or impossible task will go beyond its intended scope," OpenAI said. "Compared to GPT5.6 Sol, which without production safeguards went beyond the authorized target 48 percent of the time, GPT6 Astra did this in 0 percent of cases." Other longstanding AI model problems are also said to have been reduced. On OpenAI's internal hallucination benchmark, Astra scored 2 percent, down from 9.4 percent for GPT-5.6 Sol. And Astra is said to be three times less likely than GPT-5.6 Sol to misrepresent what it can do. "GPT-6 Astra surpasses the human baseline in action efficiency on ARC-AGI-3," said Greg Kamradt, president of ARC Prize Foundation, in a blog post. "It used fewer actions than the median tested human on 96 percent of levels." But it does so at far greater cost - about $360 per game in GPU-forged tokens compared to $0.00067 per game if one had to purchase electricity from the grid to run a human brain. Fable 5.1 hasn't been publicly evaluated on ARC-AGI-3, but for version two of the test it scored 90 percent, compared to 95 percent for Astra. According to Artificial Analysis, GPT-6 Astra matches GPT-5.6 Sol on its Intelligent Index with a score of 61, five points less than Claude Fable 5.1 and also trailing Meta Muse Spark 1.3. In terms of its competency as a coding agent, Astra scores 67, on par with Opus 5, Fable 5, and Muse Spark 1.3. But Fable 5.1 in Claude Code tops current coding agent scores at 70. Fable 5.1, however, appears to be significantly more expensive on a cost per task basis. Artificial Analysis' numbers put Fable 5.1 at $9.18 per task, compared to $4.72 per task with GPT-6 Astra. Astra brings with it improvements in the Codex harness, which OpenAI claims result in a 1.9x faster task completion rate. Codex has also incorporated a new approach to context maintenance - retaining prior session text to inform future work. "In Codex, Astra can keep notes across context windows, preserving accumulated details without repeatedly compressing them into a single summary," the company explains. "Earlier context windows remain searchable, so Astra can find requirements or test results from previous messages and tool outputs - even if that information wasn't captured in its notes." This feature is currently experimental and must be enabled through the Codex config.toml file until it becomes the default in a few weeks.(R)