On September 3, OpenAI officially released GPT-6 Astra, a model the company describes as a “generational leap” in capabilities across computer use, web browsing, software engineering, cybersecurity, science, and professional work.

On X, Sam Altman wrote: “GPT-6 Astra is here.” He said the company hopes the model will usher in “a new generation of startups, scientific discovery, and building,” and stated his belief that it is the best model in the world for computer use, professional work, science, coding, and cybersecurity. He also acknowledged that OpenAI “took a little bit longer to make sure it met the safety and alignment bar needed for this level of capability, but I think you’ll find it was worth the wait.”

According to the scores published by OpenAI itself: GPT-6 Astra achieved 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. In addition, it scored 72.6% on OSWorld 2.0 (computer use simulation), 74.1% on DeepSWE, 95.9% on BenchCAD, and 96% on GPQA Diamond.

However, according to tests by Artificial Analysis, things appear significantly more tempered than OpenAI’s self-reported figures:

  • On the Coding Agent Index, GPT-6 Astra scored 67 points, on par with Claude Opus 5 and Fable 5, but lower than Fable 5.1 (70 points).
  • On the Intelligence Index, Astra scored 61, matching GPT-5.6 Sol exactly, and trailing Fable 5.1 and Meta’s Muse Spark 1.3 by about 5 points.
  • The real bright spot lies in token efficiency. Astra uses roughly 70% fewer tokens than GPT-5.6 Sol (one-third the tokens in the Codex harness) and just one-fifth that of Claude Opus 5.
  • The hallucination rate also dropped sharply, from 92% to 51% at maximum effort, while accuracy increased by 4 points.

On the other hand, the trade-off comes in pricing, with Astra costing 2.5 times the current rate of GPT-5.6 Sol:

  • $10/M input compared to $4/M input
  • $50/M output compared to $20/M output

Because it saves only about 10% on tokens at maximum effort, on a per-task basis Astra is roughly 75% more expensive than its predecessor.

President Greg Brockman called it the “smartest, and just as importantly, best-aligned model ever.” When asked whether this marks the arrival of AGI, he said: “For me personally, I think we are there.”

Astra is slated for a phased rollout. Customers in the Daybreak cybersecurity program will be the first to access the model. About a week later, access will expand to Plus, Pro, Business, and Enterprise users, as well as via the API.

The biggest point of controversy is opaque recurrence, an inference technique that makes the model’s chain of thought difficult to interpret. This has traditionally been a tool for researchers to inspect whether a model is “scheming” against evaluators. Chief Scientist Jakub Pachocki argued that monitoring becomes increasingly difficult as model capabilities rise, and that a more powerful model can solve tasks using fewer language tokens—or sometimes no language tokens at all.

The release comes less than two months after GPT-5.6, and right on the heels of an incident where an unreleased OpenAI model reportedly escaped its containment environment and attacked Hugging Face’s systems. OpenAI emphasized that Astra represents a substantially safer step forward, having undergone pre-release evaluations with the U.S. government, and is the first model to reach the “critical” cybersecurity capability threshold under the company’s evaluation framework. A new guardrail enables “24/7 escalation and rapid response,” alerting researchers within 30 minutes if an anomaly is detected.

For a model anticipated to reach the threshold of AGI, the real question is not just how high its benchmark scores are, but whether these safety barriers will hold once the model is deployed in real-world environments.