Log inCompare

4 September 2026

OpenAI releases GPT-6 Astra, Anthropic cuts Claude cache costs 75%

First reported

Deep Learning Weekly, Sloth Bytes and 1 other ran this on , a day before the next source picked it up.

  • OpenAI released GPT-6 Astra, optimized for automating computer tasks and testing. Artificial Analysis benchmark rankings show it second after Anthropic's Claude Fable 5.1.
  • Anthropic released Claude Fable 5.1 and Mythos 5.1 with cache read costs reduced to $0.25 per million tokens, down 75% from prior pricing.
  • Astra scored differently across test harnesses: 62.7% with standard testing and 99.9% with OpenAI's custom adapter. Independent evaluator Artificial Analysis ranked it first across 50+ benchmarks with 169 points.
  • Artificial Analysis updated its Intelligence Index to version 4.2, adding real-world knowledge and PDF document tests while removing a solved benchmark and weighting private test data more heavily.

Where they differ

  • Sloth Bytes

    questioned whether benchmarks reflect real-world capability.

  • TLDR AI

    emphasized Astra's extreme benchmark score on one test harness.

  • Deep Learning Weekly

    focused on Anthropic's cost reductions. All three newsletters missed that independent evaluator Artificial Analysis ranked Astra first overall, contradicting some of their framings.

What each one reported

Sloth BytesThe Coding Sloth

OpenAI released GPT-6 Astra optimized for computer use tasks and QA testing, while Anthropic released Claude Fable 5.1 with 75% cheaper cache reads. The newsletter argues Astra excels at long-running automation jobs, Fable at ambitious coding work, and benchmarks should be taken with skepticism since Gemini 3.8 Flash ties Astra on public leaderboards.

TLDR AITLDR editorial team

GPT-6 Astra scored 62.7% on ARC-AGI-3 Semi-Private with standard harness and 99.9% with provider adapter harness. The newsletter notes it turned unfamiliar environments into compact symbolic world models using fewer actions than median human on 96% of levels.

Deep Learning WeeklyEditorial team

Anthropic released Claude Fable 5.1 and Mythos 5.1, reducing cache read costs by 75% to $0.25 per million tokens and more than doubling Terminal-Bench-Science scores to 52.6%.

Reported by The Decoder