Wikiprompt
Blog›Guides

LLM Benchmark Snapshot: Week of September 13, 2026

The weekly photo across five leaderboards: Claude Fable 5.1 holds text, LiveBench and BenchLM; GPT-5 keeps the Aider coding crown; GPT-6 Astra dominates webdev. Stability at the top, split by discipline.

LLM Benchmark Snapshot: Week of September 13, 2026

The weekly photo of the LLM leaderboards is in: snapshot 2026-09-13, across five independent sources. The headline this week is stability at the top and a three-way split by discipline.

Leaders by source

  • LMArena text: Claude Fable 5.1 (max) holds #1 at 1510 Elo, with Claude Opus 5 variants at #2 and #3 - an all-Anthropic podium.
  • LiveBench: Claude Fable 5.1 (max effort) leads at 83.78, ahead of Fable 5 and GPT-6 Astra (83.05). Same leader as last week.
  • BenchLM: Claude Fable 5.1 at 84.61 over GPT-6 Astra (84.08) - the tightest gap of the five boards.
  • Aider polyglot (coding): GPT-5 (high) keeps the coding crown at 88.0, with GPT-5 medium and o3-pro behind. Anthropic does not crack this top 3.
  • LMArena webdev: GPT-6 Astra (max) dominates at 1800 Elo, 42 points clear of Fable 5.1.

    • What changed vs last week

    Very little at the top, and that is the story: the same two families have now split the boards for two straight weeks. Anthropic owns general text and mixed-task benchmarks (text arena, LiveBench, BenchLM); OpenAI owns the developer boards (Aider coding, webdev arena). If you pick a model by leaderboard, the answer now genuinely depends on which leaderboard.

    The sources-disagree angle

    A single "best model" claim hides a 3-way split: Fable 5.1 is #1 on three boards, GPT-6 Astra on one, GPT-5 on one. That is exactly why we snapshot all five daily - each board measures a different skill, and the receipts stay on the record.

    Browse the full tables at /benchmarks, or compare over time in the history.

    Tags
    benchmarks·llm·leaderboards·claude·gpt-5·weekly-snapshot

    Related Articles