GPT‑6 Astra
Blog · Watch/Read · (gathered 2026-09-04)
- Simon highlights the release of OpenAI’s GPT-6 Astra, positioned as a direct competitor to Claude Fable with matching API pricing ($10/M input, $50/M output), though he hasn’t had personal access yet.
- Astra shows significant improvements in security tasks (100% on ExploitBench) and long-context handling (100% on OpenAI’s eight-needle benchmark up to 512K tokens).
- While achieving 99.9% on ARC-AGI 3 using a custom “Provider Adapter harness,” it scores lower than Claude Fable 5.1 on Artificial Analysis’s Intelligence Index but leads in cost efficiency on their Coding Agent Index.
- The impressive ARC-AGI score comes with a notable caveat regarding the use of a custom harness and higher cost ($19K vs $26K for default harness), underscoring the importance of scrutinizing benchmark methodologies.