Skip to content
LLM Atlas

LLM Atlas

Every other leaderboard tells you a model scored 92 on a benchmark. This one also tells you whether that benchmark still means anything, and who reported the number.

Open the leaderboardHow we judge benchmarks

Here is a real row from the leaderboard, exactly as the data has it.

Worked example

Claude Fable 5(xhigh)Anthropic

98.5%on ARC-AGI

saturatedcontaminatedsuperseded

The same model scores 88.3% on ARC-AGI-2, the successor built because the first one stopped separating models. Both numbers are real. Only one of them is worth quoting.

Who is ahead, and on what

Each card is one row from the leaderboard, picked by the rule in its heading. “Highest on a benchmark” is not “best”.

  • Highest on GPQA

    GPT-6 Astra (max)OpenAI

    95.8%ind.

    nearing saturationsupersededabout GPQA

  • Highest on Terminal-Bench

    GPT-5.5OpenAI

    84.7%ind.

    activecontaminatedabout Terminal-Bench

    A system score: the model and its harness measured together.

  • Highest Arena Elo, text to image

    gpt-image-2.5-sunburstOpenAI

    1425 ±6.8Arena

    No health flags: no benchmark health record exists for an arena board.

    A preference ranking from public pairwise votes, not a benchmark score.

Open weights, price and the full ranking

Benchmark metadata from benchwiki. Scores from Epoch AI (CC BY 4.0). Pricing from vendor pages where readable, and from OpenRouter (CC BY 4.0) otherwise. Arena Elo from Arena (CC BY 4.0). Benchmark data fetched 2026-10-18.