Key Points
- OpenAI's GPT-6 Astra scores 99.9% on ARC-AGI-3, effectively saturating a benchmark designed to test genuine creative reasoning on novel visual puzzles.
- The model claims world-leading computer use capabilities, a practical differentiator for commercial adoption that may matter more than reasoning benchmarks.
- ARC-AGI's creators are already developing a fourth benchmark version, signaling saturation on v3 was an anticipated milestone as frontier models improve.
Summary
OpenAI's GPT-6 Astra Saturates ARC-AGI-3 Benchmark
OpenAI has released GPT-6 Astra, scoring 99.9% on ARC-AGI-3—effectively saturating one of the field's most rigorous reasoning benchmarks. The model also demonstrates strong performance across Terminal Bench science tasks and computer use, with OpenAI describing it as the world's best computer use model.
The ARC-AGI-3 result is significant because the benchmark is designed to require genuine creative reasoning. Unlike narrow task optimization, ARC-AGI tests require understanding novel visual puzzles and applying learned logic to new problems. A 99.9% score leaves virtually no room for improvement on that particular test.
Beyond reasoning, Astra scored 64.6% on Terminal Bench science tasks (with a cost of $26.26 per evaluation) and showed strong performance on the Exploit Jim Honeypot benchmark, where lower scores are better. The computer use capability stands out as a potential practical differentiator—if the claims hold, it could matter more for commercial adoption than reasoning benchmarks.
The team behind ARC-AGI is already working on a fourth version of the benchmark, signaling that saturation on v3 was always a likely outcome as frontier models improved. Sam Altman is scheduled to discuss the launch on Bloomberg.
The rollout was notably chaotic, with the blog post going live on OpenAI's website before official social media announcement, limiting immediate public reaction and analysis.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.