LLM Benchmarks 2019–2026: The Records on 16 Leaderboards

Epoch Capabilities Index

published 2025 · models from 2023 to 2026 · 274 models in the archive · active · results: Epoch AI

Combines the results of more than 50 benchmarks into one number per model, so that models can be compared even if they were never tested on the same benchmarks. It is an index, not a percentage: Epoch AI scales it so that Claude 3.5 Sonnet stands at 130 and GPT-5 at 150.

Leaders today: Claude Opus 5.5 (167.3) · GPT-6 Astra (166.4) · GPT-6.1 Sol (166.1) · Claude Sonnet 5.5 (165.0) · Claude Fable 5.1 (164.7)

METR Time Horizon

published 2025 · models from 2019 to 2026 · 39 models in the archive · active · results: METR

The length of the tasks a model completes correctly about half of the time, measured by how long a human expert needs for them. The number is a duration, not a score: a longer horizon means the model can work on its own for longer.

The bars use a logarithmic scale: each step to the right is a multiplication.

Leaders today: Claude Mythos Preview (Early) (17 h 25 min) · Claude Opus 4.6 (unknown thinking) (11 h 59 min) · Gemini 3.1 Pro Preview (6 h 24 min) · GPT-5.2 (high) (5 h 52 min) · GPT-5.3 Codex (5 h 50 min)

MMLU

published 2020 · models from 2021 to 2024 · 158 models in the archive · no new results since 2024 · results: papers and Stanford HELM, collected by Epoch AI

Multiple-choice questions from 57 subjects, from high-school level to professional exams. For years the standard test of general knowledge; today the best models are so close to the ceiling that hardly any new results are published.

GPT-3's well-known 43.9% is filed in the archive under the API version of January 2022 and therefore does not appear as a record here.

Leaders today: GPT-4o (88.1%) · Claude 3.5 Sonnet (20241022) (87.3%) · DeepSeek v3 (87.2%) · Gemini 1.5 Pro (002) (86.9%) · Claude 3.5 Sonnet (20240620) (86.5%)

GPQA Diamond

published 2023 · models from 2023 to 2026 · 214 models in the archive · active · results: Epoch AI (own runs)

198 multiple-choice questions in biology, physics and chemistry at graduate level, written so that non-experts fail even with web access. Guessing gives 25%.

Leaders today: GPT-6 Astra (max) (95.8%) · Claude Sonnet 5.5 (max) (95.6%) · Gemini 3.8 Flash (high) (95.4%) · GPT-6.1 Sol (max) (95.4%) · Gemini 3.7 Flash (high) (94.8%)

Humanity's Last Exam

published 2025 · models from 2024 to 2026 · 44 models in the archive · active · results: Humanity's Last Exam leaderboard, collected by Epoch AI

Expert-written questions at the edge of human knowledge across many fields, built to stay hard after older tests were solved.

Leaders today: GPT 6 Astra (54.8%) · Fable 5.1 (xhigh) (46.5%) · Gemini 3.1 Pro Preview (thinking high) (46.4%) · Gemini 3.8 Flash (44.5%) · GPT-5.4 Pro (44.3%)

SimpleQA Verified

published 2025 · models from 2024 to 2026 · 79 models in the archive · active · results: Epoch AI (own runs)

Short factual questions with exactly one correct answer, answered without tools. It shows how much a model really knows and how often it invents an answer instead of admitting that it does not know.

Leaders today: GPT-6 Astra (max) (75.6%) · GPT-6.1 Sol (max) (73.9%) · Gemini 3.1 Pro Preview (73.5%) · Claude Opus 5.5 (max) (72.2%) · Claude Fable 5.1 (max) (70.8%)

GSM8K

published 2021 · models from 2020 to 2024 · 146 models in the archive · no new results since 2024 · results: papers and Stanford HELM, collected by Epoch AI

Math word problems at grade-school level that need several calculation steps. Once a hard test of reasoning, now practically solved.

Leaders today: DS-Coder-V2-Instruct (94.5%) · Qwen2.5-Coder-14B-Instruct (94.2%) · Qwen2.5-Coder-32B-Instruct (93.0%) · GPT-4 (92.0%) · GPT-4 (92.0%)

MATH Level 5

published 2021 · models from 2023 to 2025 · 94 models in the archive · active · results: Epoch AI (own runs)

The hardest level of the MATH dataset: competition problems from algebra, geometry, number theory and probability.

Leaders today: GPT-5 (high) (98.1%) · o4-mini (high) (97.8%) · o3 (high) (97.8%) · GPT-5 mini (high) (97.8%) · Claude Sonnet 4.5 (32k thinking) (97.7%)

FrontierMath (Tiers 1–3)

published 2026 · models from 2024 to 2026 · 83 models in the archive · active · results: Epoch AI (own runs)

Original, unpublished mathematics problems written by mathematicians, from advanced university level up to research level. The problems are kept private so that models cannot have seen them in training.

The archive lists this second version of the problem set since June 2026, with results for older models as well.

Leaders today: GPT-6 Astra (max) (93.7%) · GPT-6.1 Sol (max) (93.7%) · Claude Opus 5.5 (max) (91.2%) · Claude Fable 5.1 (max) (90.2%) · GPT-6 Sol (max) (89.8%)

SWE-bench Verified

published 2024 · models from 2024 to 2026 · 32 models in the archive · active · results: Epoch AI (own runs)

500 real issues from open-source Python projects, checked by humans. The model has to change the code so that the project's tests pass.

These are Epoch AI's own runs with one fixed setup; vendors often report higher numbers with their own tooling.

Leaders today: Claude Opus 4.7 (max) (83.5%) · GPT-5.5 (xhigh) (80.6%) · Gemini 3.5 Flash (79.3%) · Claude Opus 4.6 (no thinking) (78.7%) · GLM-5.2 (78.7%)

Aider Polyglot

published 2024 · models from 2024 to 2025 · 54 models in the archive · active · results: Aider leaderboard, collected by Epoch AI

225 difficult programming exercises in C++, Go, Java, JavaScript, Python and Rust. It measures whether a model can edit existing code correctly, with a second attempt after seeing the test output.

Leaders today: GPT-5 (high) (88.0%) · o3-pro (84.9%) · Gemini 2.5 Pro Preview (Jun 2025) (83.1%) · o3 (high) (81.3%) · Grok 4 (79.6%)

Terminal-Bench

published 2025 · models from 2025 to 2026 · 45 models in the archive · active · results: Terminal-Bench leaderboard, collected by Epoch AI

Tasks that an agent has to solve in a real command line: set up software, fix a system, process data. The score belongs to a model together with its agent software; we show the best result per model.

Leaders today: GPT-5.5 (84.7%) · GPT-5.4 (81.8%) · Gemini 3.1 Pro (80.2%) · Claude Opus 4.7 (80.2%) · Claude Opus 4.6 (79.8%)

ARC-AGI

published 2019 · models from 2024 to 2026 · 91 models in the archive · active · results: ARC Prize leaderboard, collected by Epoch AI

Grid puzzles in which a rule has to be worked out from a few examples and applied to a new case. Easy for people, long out of reach for language models — the test of learning something new instead of recalling it.

Leaders today: Claude Fable 5 (XHigh) (98.5%) · Gemini 3.8 Flash (High) (98.5%) · GPT-6 Astra (XHigh) (98.5%) · Claude Opus 5.5 (High) (98.5%) · GPT-6.1 Sol (XHigh) (98.5%)

ARC-AGI-2

published 2025 · models from 2024 to 2026 · 86 models in the archive · active · results: ARC Prize leaderboard, collected by Epoch AI

The successor to ARC-AGI from 2025, with puzzles chosen to resist the methods that solved the first version.

Leaders today: GPT-6 Astra (Max) (95.0%) · GPT-6.1 Sol (Max) (94.2%) · Claude Opus 5.5 (High) (93.3%) · GPT-5.6 Sol (Max) (92.5%) · Claude Opus 5 (Max) (90.4%)

HellaSwag

published 2019 · models from 2019 to 2024 · 97 models in the archive · no new results since 2024 · results: papers and Stanford HELM, collected by Epoch AI

Pick the most plausible continuation of an everyday scene out of four. A test of common sense from 2019 that people find trivial; guessing gives 25%.

Leaders today: GPT-4 (95.3%) · LLaMA-3.1 405B Base (89.2%) · Falcon-180B (89.0%) · DeepSeek-V3 Base (88.9%) · DeepSeek-V2 Base (87.1%)

WinoGrande

published 2019 · models from 2019 to 2024 · 84 models in the archive · no new results since 2024 · results: papers, collected by Epoch AI

Decide what an ambiguous word in a sentence refers to — possible only with an understanding of the situation. Guessing gives 50%.

Leaders today: LLaMA-3.1 405B Base (89.2%) · Claude 3 Opus (88.5%) · GPT-4 (87.5%) · Falcon-180B (87.1%) · UNICORN (86.6%)