GPT-6 Astra is OpenAI’s first Stargate and lightly recurrent supermodel, [released today](http://GPT-6 Astra’s launch). It cleanly outperforms Fable 5.1 on many benchmarks, including fully saturating the hardest version of FrontierMath (97.6%) and ARC-AGI-3 (99.9%). Many demos will focus on the usual promotional highlights, such as computer use, playing Pokémon, Blender, and science and cybersecurity benchmarks (the system card). Greg says that AGI has arrived, while Jakub says this is finally the automated AI research intern he has always wanted.
We are not qualified to comment on these claims, but we were granted early access and put it to work on every practical, real-world task we could think of. After consuming more than 20 billion Astra tokens, we can confirm the most surprising finding: GPT-6 Astra belongs to an entirely new class of models[1]—models that are natively capable of performing the full range of an AI engineer’s work. They can now help you select and train models, label data (both assisting with annotation and using your labels to perform active learning, such as SAM), keep pipelines running at full capacity, instrument systems and read logs, deploy and debug complete systems in one shot, and scale up tasks while directing and evaluating sub-agents (including agents that run other models), all while maintaining coherence across billions of tokens in a single agent thread.
Raise Your Ambitions
We previously wrote about the high-return activity of raising your expectations of LLMs. Our experience has made us bolder than ever, and our ambitions have risen exponentially. Over the past month, we went from launching a fun “end my SaaS” competition and soliciting ideas from humans[2] to building more than a dozen interns…