Benchmarks

The benchmarks Tokscale tracks, grouped by what they test. Each one keeps its own unit, source and license. Tokscale does not combine them into one score.

12 benchmarks with results, 1 linked onlyFrom Epoch AI, LMArena, Terminal-BenchRetrieved Oct 21, 2026

Overall

1 benchmark
  • Epoch Capabilities Index

    Combines scores from many benchmarks into one general capability scale, so models can be compared even after single benchmarks saturate.

    By Epoch AI

    Unit
    ECI
    Models
    160 matched270 in the source
    Last update
    Oct 15, 2026Newest model releaseRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Agentic coding

5 benchmarks
  • Terminal-Bench 4.0

    Agents work on tasks inside a terminal. The score is the share of tasks resolved, with a 95% confidence interval.

    By Terminal-Bench

    Unit
    % resolved
    Models
    23 matched23 in the source
    Last update
    Oct 6, 2026Newest publish dateRetrieved Oct 21, 2026
    License
    Terms not stated
  • SWE-Bench Pro V2Link only

    Long-horizon software engineering tasks in public open-source repositories. V2 is a refreshed public split of 642 tasks with a locked evaluation protocol (Scale AI, September 22, 2026).

    By Scale AI

    Results not republished. Open the leaderboard on Scale AI's site.

    Open the SWE-Bench Pro V2 leaderboard (opens in a new tab)
    Unit
    % resolved
    Models
    Not republishedResults stay on the source site
    Last update
    —See the leaderboard
    License
    Leaderboard: all rights reserved (Scale AI)
  • CursorBench

    Cursor evaluates agents on ambiguous, multi-file tasks taken from real Cursor sessions.

    By Cursor · collected by Epoch AI

    Unit
    score
    Models
    18 matched18 in the source
    Last update
    Oct 12, 2026Newest model releaseRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)
  • DeepSWE

    Coding agents on original, long-horizon software engineering tasks. The score is pass@1.

    By Datacurve · collected by Epoch AI

    Unit
    pass@1
    Models
    26 matched26 in the source
    Last update
    Sep 3, 2026Newest model releaseRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)
  • FrontierSWE

    Very long software engineering tasks at the edge of human ability. The score is the mean of five trials per task.

    By FrontierSWE · collected by Epoch AI

    Unit
    mean@5
    Models
    21 matched21 in the source
    Last update
    Oct 7, 2026Newest model releaseRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Web development

1 benchmark
  • WebDev Arena

    Human preference for web development: people compare what two anonymous models build for the same prompt and vote. Scores are Bradley-Terry ratings with 95% confidence intervals.

    By LMArena

    Unit
    Arena score
    Models
    110 matched129 in the source
    Last update
    Oct 19, 2026Newest publish dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Agents

1 benchmark
  • Agent Arena

    Signals from real agent sessions on Arena (task success, user feedback, tool errors) turned into an estimated net improvement per orchestrator model, with 95% confidence intervals.

    By LMArena

    Unit
    net improvement
    Models
    48 matched51 in the source
    Last update
    Oct 19, 2026Newest publish dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Reasoning

3 benchmarks
  • GPQA Diamond

    Hard multiple-choice questions in biology, chemistry and physics, written by PhD-level experts.

    By Epoch AI

    Near its ceiling: the top score is 96.4%, and Epoch AI estimates the reachable maximum at 97%. Differences between top models are small.

    Unit
    % correct
    Models
    151 matched224 in the source
    Last update
    Oct 15, 2026Newest run dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)
  • Humanity's Last Exam

    2,500 expert-written questions across more than 100 academic subjects that need deep, specialised knowledge.

    By Center for AI Safety and Scale AI · collected by Epoch AI

    Unit
    % correct
    Models
    40 matched46 in the source
    Last update
    Oct 7, 2026Newest model releaseRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)
  • FrontierMath Tiers 1–3 (v2)

    Expert-written math problems from advanced undergraduate level to early career research.

    By Epoch AI

    Near its ceiling: the top score is 93.7%, and Epoch AI estimates the reachable maximum at 95%. Differences between top models are small.

    Unit
    % solved
    Models
    86 matched87 in the source
    Last update
    Oct 15, 2026Newest run dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Knowledge

1 benchmark
  • SimpleQA Verified

    1,000 short factual questions about politics, science and technology, art, sports, geography, music and more.

    By Epoch AI

    Unit
    % correct
    Models
    80 matched82 in the source
    Last update
    Oct 15, 2026Newest run dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

History

1 benchmark
  • SWE-bench VerifiedHistory

    GitHub issues from real Python repositories: can the model write a valid fix?

    By Epoch AI

    History only. Newer agentic coding benchmarks are on the main board.

    Unit
    % resolved
    Models
    32 matched32 in the source
    Last update
    Jun 25, 2026Newest run dateRetrieved Oct 21, 2026
    License
    CC BY 4.0 (opens in a new tab)

Sources and attribution

Each score keeps its own source, unit and run conditions. Tokscale does not combine benchmarks into one score, and a missing score means the model was not evaluated, not that it scored zero.

  • Epoch AIEpoch Capabilities Index, Epoch's own runs and the external results it collects.CC BY 4.0 · Retrieved Oct 21, 2026 · Data through Oct 15, 2026Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from https://epoch.ai/benchmarks.
  • LMArenaWebDev Arena and Agent Arena leaderboards.CC BY 4.0 · Retrieved Oct 21, 2026 · Data through Oct 19, 2026LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), Hugging Face.
  • Terminal-BenchTerminal-Bench 4.0 official leaderboard rows (results only, no task content).Terms not stated · Retrieved Oct 21, 2026 · Data through Oct 6, 2026Terminal-Bench 4.0 leaderboard, tbench.ai. Every score links back to the leaderboard.
  • TokscaleActive users, tokens and token mix from public Tokscale submissions. Hidden accounts are excluded.How we count