Summary & Key Takeaways ​

• GPT-6 Astra is rolling out to users and APIs, priced competitively with Claude Fable 5/5.1.
• It shows impressive scores on ARC-AGI 3 (99.9% with custom harness) and security benchmarks like ExploitBench (100%).
• Astra demonstrates strong long-context capabilities, achieving high accuracy up to 1M tokens.
• Despite its strengths, it trails Claude Fable 5.1 and Meta's Muse Spark 1.3 on Artificial Analysis's Intelligence Index.
• It leads the Coding Agent Index for cost efficiency, indicating strong performance in code-related tasks.

Our Commentary ​

Simon Willison's analysis of GPT-6 Astra is exactly what we needed after the official announcement. He cuts through the marketing to give us the real picture: the benchmarks, the pricing, and where it truly stands against competitors. The nuance around ARC-AGI scores and the security prowess are particularly telling. This is a serious contender, but not an undisputed champion.