Skip to content
little agentclub.

Models, compared.

Pick a task. See independently tested scores.

Answer hard science questions.

Share of expert-level questions answered correctly. GPQA Diamond.

Independently tested by Epoch AI (opens in a new tab) · Checked

  • Highest shown

    Gemini4 Argon

    Reasoning: High

    96.4%

  • GPT6.1 Sol

    Reasoning: Max

    95.4%

  • Grok4.7

    Reasoning: Extra high

    92.7%

  • GLM5.3

    Reasoning: Max

    90.9%

  • ClaudeOpus 5.5

    Reasoning: Max

    90.6%

  • DeepSeekV4.1 Flash

    Reasoning: Max

    89.8%

Small gaps may be noise. These tests don’t measure everyday agent use.

Scores and sources

Six selected model versions, tested by Epoch AI rather than their makers. We use only Epoch’s own evaluations, not the maker-reported results it also collects. The club has not repeated these tests.

“Highest shown” means the largest published mean in this selection, not a proven winner. Scores can vary between runs. Different reasoning settings use different budgets, so this is not a comparison at equal cost.

Each percentage keeps its original meaning. We do not average the tests or add their results to Club Rank. A missing result is left blank.

We show the latest dated run for each exact configuration in the 19 October download, using its mean score rather than picking its best result. This is a dated selection, not a complete or live leaderboard. Coding tests remain in the guide below until these versions have comparable results.

GPQA Diamond

Original GPQA Diamond. Not GPQA Diamond Clean. Epoch uses Inspect, a reasoning prompt and API-default temperature. Answer choices are shuffled.

Test method (opens in a new tab)
  • Gemini 4 Argon: · 96.4% ± 1.2 points
  • Grok 4.7: · 92.7% ± 1.6 points
  • GPT 6.1 Sol: · 95.4% ± 1.4 points
  • Claude Opus 5.5: · 90.6% ± 1.9 points
  • DeepSeek V4.1 Flash: · 89.8% ± 1.7 points
  • GLM 5.3: · 90.9% ± 1.6 points

OTIS Mock AIME 2024–2025

OTIS practice problems from 2024–2025. Not the official AIME exam or AIME 2026. Epoch uses Inspect with a reasoning prompt and API-default temperature. Final answers are model-graded. A perfect observed score does not establish perfect maths ability.

Test method (opens in a new tab)
  • Gemini 4 Argon: · 100.0% ± 0.0 points
  • Grok 4.7: · 98.1% ± 1.4 points
  • GPT 6.1 Sol: · 100.0% ± 0.0 points
  • Claude Opus 5.5: · 100.0% ± 0.0 points
  • DeepSeek V4.1 Flash: · 98.3% ± 0.6 points
  • GLM 5.3: · 91.1% ± 3.5 points

SimpleQA Verified

SimpleQA Verified, not original SimpleQA. This is a factual-answer test, not a live web-search test. Epoch’s internal evaluation. Different reasoning settings can use different amounts of compute; this is not a cost-matched comparison.

Test method (opens in a new tab)
  • Gemini 4 Argon: · 76.0% ± 1.4 points
  • Grok 4.7: · 56.0% ± 1.6 points
  • GPT 6.1 Sol: · 73.9% ± 1.4 points
  • Claude Opus 5.5: · 72.2% ± 1.4 points
  • DeepSeek V4.1 Flash: · 42.9% ± 1.6 points
  • GLM 5.3: · 41.0% ± 1.6 points

± is one standard error reported by Epoch, not a 95% confidence interval. Overlapping ranges do not establish a clear winner. A zero error estimate or perfect observed score does not remove all uncertainty.

Epoch’s methods and limits (opens in a new tab) · Download scores and run records

Data adapted from Epoch AI, “Capabilities & benchmarking”, accessed 19 Oct 2026, under CC BY 4.0. Family illustrations are unofficial; the scores belong to the exact model versions and reasoning settings named above. They are not scores for Gemini, Grok Bot or other agent apps.

Explore more benchmarks

More tests for different jobs. These results open on the benchmark’s own site.

Use tools and follow rules

Explore customer-service tasks that combine conversation, tool calls and company policies.

Successful tasksτ-bench (opens in a new tab)
8 more benchmarks from model releases

Names you’ll see in launch reports. Some sources publish leaderboards; others explain the test. A maker’s reported result is not automatically an independent evaluation.

Reasoning & knowledge

  • GPQA Diamond

    Hard science questions in biology, chemistry and physics.

    Compare the Diamond subset with the same tools and reasoning settings. A high score is not a measure of everyday usefulness.

    Original benchmark (opens in a new tab)

  • AIME

    Competition maths problems, often used in model launch reports.

    Match the exam year, attempts and tool access. MAA runs the exam, not an AI leaderboard; model scores come from the evaluator’s report.

    About the exam (opens in a new tab)

  • Humanity’s Last Exam

    Difficult academic questions across many subjects, including images.

    Keep the original HLE and HLE-Diamond separate. Check the question set and whether browsing or other tools were allowed.

    Test and results (opens in a new tab)

  • ARC-AGI

    Learning unfamiliar rules, from visual puzzles to interactive tasks.

    Choose one edition. ARC-AGI-3 is interactive; earlier editions use static tasks. Check the system, cost and whether a result is verified.

    ARC Prize leaderboard (opens in a new tab)

Coding & tools

  • LiveCodeBench

    Coding problems collected from programming contests over time.

    Match the problem date range and task type. A newer test window is a different comparison, not a change in the same score.

    Test and leaderboards (opens in a new tab)

  • Terminal-Bench

    Tasks completed by an agent in a command-line environment.

    Match the benchmark edition, agent setup and resource budget. This tests a model working with tools, not the model alone.

    Terminal-Bench leaderboard (opens in a new tab)

Images & research

  • MMMU-Pro

    Understanding images, diagrams and subject-specific questions together.

    Use the Pro results. Original MMMU, its validation and test splits, and Pro are different comparisons.

    MMMU team’s results (opens in a new tab)

  • BrowseComp

    Finding hard-to-locate facts by searching across the web.

    Check browsing access and the time or reasoning budget. OpenAI’s research page describes the test; it is not a continuously updated leaderboard.

    Research and evaluation (opens in a new tab)

The model is only part of the story.

An agent adds tools, memory and instructions. A benchmark result belongs to the exact model version and setup that was tested.

Club Rank follows attention around personal agents. These tests explore what models and tested systems can do.

How to read the results

Use the test closest to your task. Compare results within that test, then check the version, setup and date.

We keep the original scales separate. A missing result means it has not been established here, not a score of zero. There is no combined model score and these results add no points to Club Rank.

Arena WebDev

WebDev · compare within the same category

Keep the task category and model settings the same. Read the vote counts and uncertainty ranges alongside the score; nearby places can overlap.

A preferred website is not a security, accessibility or production-readiness audit.

How Arena works (opens in a new tab)

SWE-bench

Verified · Bash Only

The default Verified Bash Only view uses the same mini-SWE-agent environment across models. Keep the benchmark subset, model version and setup fixed when comparing results.

A score on existing issues does not tell you how well a model will design a new app. Verified, Lite and Multilingual are different tests.

SWE-bench evaluation documentation (opens in a new tab)

OSWorld 2.0

OSWorld 2.0 · computer use

Check the task set, model settings and action budget. Full completion and partial credit answer different questions. Keep OSWorld 1.0 and 2.0 results separate.

These results cover the tested model and computer-use setup. They are not scores for every consumer app using that model.

OSWorld 1.0 and version notes (opens in a new tab)

τ-bench

Choose one version, domain and text or voice mode

Compare within one leaderboard. Text and voice, banking and retail, and different benchmark versions are separate conditions. Pass¹ and repeated-success measures are not interchangeable.

Reliable performance in a simulated support task does not prove that an agent is ready to manage your personal accounts.

τ-bench code and evaluation methods (opens in a new tab)

Sources checked . The club has not run these evaluations.

Meet the models