GPT‑6 Astra
Simon Willison’s Weblog Simon Willison ● Covered by 5 sources
OpenAI is rolling out GPT-6 Astra to some users today, then wider in the next few days. It’s priced like Claude Fable and looks strong on security and long-context tests.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI says GPT-6 Astra is starting with a limited set of organizations today, then opening up over the next few days to ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS. The API name will be gpt-6-astra once it lands. Simon Willison hasn’t used it yet, so the early read comes from OpenAI’s own numbers and a few outside comparisons.
The pricing is set at $10 per million input tokens and $50 per million output tokens, the same rate Willison cites for Claude Fable 5 and 5.1. That makes the model look like a direct shot at Anthropic rather than a broad, all-purpose release. And on OpenAI’s self-reported benchmarks, Astra appears to come out ahead of Fable on most of them.
The flashiest result is a 99.9% score on the ARC-AGI 3 benchmark, which was released in March. But there’s a catch. The ARC-AGI blog says that result came from OpenAI’s custom Provider Adapter harness and cost $19K. Using the default ARC-AGI harness, the score was 62.7% at a cost of $26K. That harness keeps opaque reasoning state between requests and uses compaction for longer conversations, so the model can reuse earlier work. In other words, the headline number is real, but it is not the whole story.
Astra also looks unusually strong on security work. It scored 100% on ExploitBench, compared with GPT-5.6 Sol’s 78.5%, 42.4% on ExploitGym versus Sol’s 30.3%, and 99.2% within four attempts on SRE-Bench binary reverse engineering, where Sol got 68.7%. On OpenAI’s eight-needle benchmark, it hit 100% at 256K to 512K tokens and 96.3% at 512K to 1M tokens. That suggests OpenAI has at least partially solved one of the messiest long-context problems.
Still, Astra doesn’t cleanly dominate everything. Artificial Analysis puts it at 61 on its Intelligence Index, tied with GPT-5.6 Sol and five points behind Claude Fable 5.1, while also trailing Meta’s Muse Spark 1.3. On the Coding Agent Index, though, Astra looks better: at max effort it costs about the same as GPT-5.6 Sol while scoring two points higher, and per task it comes in at less than half the cost of Claude Fable 5 for the same score.
My take — AI-written commentary, not fact-checked reporting
This looks less like a triumphal model launch than a spreadsheet war with very expensive fireworks. The real tell is the custom harness: when a benchmark needs special plumbing to show its best face, the marketing team is doing half the work. OpenAI may still have the stronger practical product here, but the industry keeps rewarding whichever lab can make the scorecard look cleanest.
Read more about this at: Simon Willison’s Weblog