Acropolis

Acropolis: AI benchmark results

01

The frontier today

Data as of
  • Sources tracked

    7

    7 current

    As of
  • Benchmarks tracked

    53

    52 with results

    As of
  • Models tracked

    930

    Canonical registry

    As of
  • Results

    32,061

    28,324 rows

    As of

Consensus capability leaders

All models

Bradley-Terry latent capability score (0 to 100) across shared evaluation boards.

  • Grade A. Evidence grade A of A to D: strongest.
    18 boards: 100.0 (scale 0.0 to 100.0), 95% confidence interval 93.4 to 100.018 boards100.0(93.4–100.0)
  • #2Fable 5.2Anthropic
    Grade A. Evidence grade A of A to D: strongest.
    15 boards: 95.3 (scale 0.0 to 100.0), 95% confidence interval 83.7 to 100.015 boards95.3(83.7–100.0)
  • Grade A. Evidence grade A of A to D: strongest.
    19 boards: 92.8 (scale 0.0 to 100.0), 95% confidence interval 88.1 to 95.519 boards92.8(88.1–95.5)
  • #4Opus 5.5Anthropic
    Grade A. Evidence grade A of A to D: strongest.
    23 boards: 92.4 (scale 0.0 to 100.0), 95% confidence interval 84.8 to 99.023 boards92.4(84.8–99.0)
  • #5Sonnet 5.5Anthropic
    Grade A. Evidence grade A of A to D: strongest.
    22 boards: 91.6 (scale 0.0 to 100.0), 95% confidence interval 83.7 to 96.822 boards91.6(83.7–96.8)
  • Grade A. Evidence grade A of A to D: strongest.
    15 boards: 91.1 (scale 0.0 to 100.0), 95% confidence interval 84.1 to 97.315 boards91.1(84.1–97.3)

Top row per benchmark

See all 37 benchmarks
  • LMArena Text

    Independent. Run by an operator independent of the model vendors.Tied #1 (2)

    LMArena · version 1 · ranked by arena rating

    Gemini 4 Carbon · high (Google)

    1,529 to 1,548

    4273 runs · as of 17 Nov 2026

  • GPQA diamond

    Independent. Run by an operator independent of the model vendors.Tied #1 (35)

    Epoch AI Benchmarking Hub · version 1 · ranked by accuracy (best score across scorers)

    Gemini 4 Carbon · high (Google)

    96.5%94.2% to 98.9%

    Run count not stated · as of 17 Nov 2026

  • CursorBench

    Vendor-run. Run by the vendor of at least one model on the list.

    CursorBench · version 4.0 · ranked by score

    Gemini 4 Carbon · high (Google)

    60.4%

    Run count not stated · as of 17 Nov 2026

  • Vals Index

    Independent. Run by an operator independent of the model vendors.Tied #1 (5)

    Vals AI benchmarks · version 2.2 · ranked by accuracy

    Fable 5.2 · high (Anthropic)

    71.3%69.4% to 73.2%

    Run count not stated · as of 17 Nov 2026

  • LMArena Text (style control)

    Independent. Run by an operator independent of the model vendors.Tied #1 (2)

    LMArena · version 1 · ranked by arena rating

    Gemini 4 Carbon · high (Google)

    1,517 to 1,536

    4273 runs · as of 17 Nov 2026

  • ARC-AGI-2

    Independent. Run by an operator independent of the model vendors.

    Epoch AI Benchmarking Hub · version 1 · ranked by score

    GPT-6 Astra · max (OpenAI)

    95%

    Run count not stated · as of 17 Nov 2026

  • Vibe Code Bench

    Independent. Run by an operator independent of the model vendors.Tied #1 (7)

    Vals AI benchmarks · version 1.1 · ranked by accuracy

    Fable 5.2 · high (Anthropic)

    93.8%91.5% to 96%

    Run count not stated · as of 17 Nov 2026

  • LMArena Vision

    Independent. Run by an operator independent of the model vendors.Tied #1 (6)

    LMArena · version 1 · ranked by arena rating

    Fable 5 · high (Anthropic)

    1,319 to 1,332

    15536 runs · as of 17 Nov 2026

  • Terminal Bench

    Independent. Run by an operator independent of the model vendors.Tied #1 (8)

    Epoch AI Benchmarking Hub · version 2.0 · ranked by accuracy

    GPT-5.5 (OpenAI)

    84.7%80.6% to 88.8%

    Run count not stated · as of 17 Nov 2026

  • Finance Agent

    Independent. Run by an operator independent of the model vendors.Tied #1 (15)

    Vals AI benchmarks · version 2.1 · ranked by accuracy

    Fable 5.2 · high (Anthropic)

    69.5%65.6% to 73.3%

    Run count not stated · as of 17 Nov 2026

  • LMArena Vision (style control)

    Independent. Run by an operator independent of the model vendors.Tied #1 (3)

    LMArena · version 1 · ranked by arena rating

    Fable 5 · high (Anthropic)

    1,303 to 1,316

    15536 runs · as of 17 Nov 2026

  • FrontierMath-2025-02-28-Private

    Independent. Run by an operator independent of the model vendors.Tied #1 (6)

    Epoch AI Benchmarking Hub · version 2025-02-28-Private · ranked by accuracy (best score across scorers)

    GPT-5.5 Pro Pre-release · high (OpenAI)

    52.4%46.7% to 58.1%

    Run count not stated · as of 17 Nov 2026

02

Recent activity

  • Evaluation

    Opus 5.5 leads FrontierCode

    54.6% on Main score (Epoch AI Benchmarking Hub, 51 evaluated rows).

    Epoch AI Benchmarking Hub

  • Evaluation

    GPT-5.2 leads GDPval

    49.7% on Win rate (Epoch AI Benchmarking Hub, 11 evaluated rows).

    Epoch AI Benchmarking Hub

  • Evaluation

    GPT-6 Astra leads HLE

    54.8% on Accuracy (Epoch AI Benchmarking Hub, 56 evaluated rows).

    Epoch AI Benchmarking Hub