LMArena · GPQA · SWE-bench · HLE · ARC-AGI · OSWorld · BFCL
AI Model Benchmarks
3086 scores for 1079 models on 23 public benchmarks. Every number links to where it came from, and self-reported vendor numbers are labelled separately from official leaderboards and independent runs.
3086 scores 1079 models 23 benchmarks Updated Oct 22, 2026
Current champions
The top model on each core benchmark. Only official leaderboards, independent evaluations and compiled data count; vendor self-reports are excluded.
LMArena Text Gemini 4 Argon 1525 Human preference GPQA Diamond Gemini 4 Argon 96.4% Science Humanity's Last Exam GPT-6 Astra 54.8% Knowledge AIME (OTIS Mock AIME 2024–2025) Gemini 4 Argon 100.0% Maths SWE-bench Verified Claude Opus 4.7 83.5% Agentic coding Terminal-Bench 2.0 GPT-5.5 84.7% Agentic coding OSWorld (Verified) Claude Sonnet 4.6 72.1% Computer use τ²-bench Gemini 3.0 Pro 85.3% Tool use & agents BFCL v4 (function calling) Claude Opus 4.5 77.5% Tool use & agents MMLU-Pro Gemini 3.1 Pro 91.2% Knowledge
Model × benchmark matrix
Each cell is the best score we have for the model on the benchmark's main setting. Official leaderboards and independent runs are preferred; a cell marked ◇ only has a vendor-reported or paper number. Click a column header to sort, tick models to compare.
| Model | LMArena Text | GPQA Diamond | Humanity's Last Exam | AIME (OTIS Mock AIME 2024–2025) | SWE-bench Verified | Terminal-Bench 2.0 | OSWorld (Verified) | τ²-bench | BFCL v4 (function calling) | MMLU-Pro |
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.2OpenAI | 1437 | 91.4% | 27.8% | 96.1% | 73.8% | 64.9% | — | 81.6% | 55.9% | 87.4% ◇ |
| Gemini 3 ProGoogle | 1486 | 92.6% | 37.5% | 91.4% | 77.4% | 69.4% | — | 75.9% | 72.5% | 90.1% ◇ |
| Claude Sonnet 4.5Anthropic | 1456 | 82.3% | 13.7% | 77.8% | 74.8% | 46.5% | 62.9% | 72.4% | 73.2% | — |
| GPT-5OpenAI | 1435 | 86.2% | 25.3% | 91.4% | 74.4% | 49.6% | — | 81.6% | — | 87.1% ◇ |
| o3OpenAI | 1432 | 81.8% | 20.3% | 84.4% | 62.3% | — | 23.0% | 73.9% ◇ | 63.0% | — |
| Claude Opus 4.5Anthropic | 1474 | 86.0% | 25.2% | 86.1% | 79.2% | 63.1% | 66.3% ◇ | 79.6% | 77.5% | — |
| Gemini 2.5 Pro (Jun 2025)Google | 1446 | 85.3% | 21.6% | 84.7% | 57.6% | 32.6% | — | — | — | 86.0% ◇ |
| Claude Sonnet 4Anthropic | 1402 | 79.2% | 7.8% | 68.9% | 76.8% | — | 43.9% | 80.5% ◇ | — | — |
| o4-miniOpenAI | 1391 | 79.6% | 18.1% | 81.7% | 64.6% | — | — | 68.3% | 53.2% | — |
| GPT-4.1OpenAI | 1415 | 66.9% | 5.4% | 38.3% | 48.5% | — | — | 74.0% | 54.0% | 81.8% ◇ |
| Claude 3.7 SonnetAnthropic | 1388 | 79.7% | 8.0% | 57.8% | 66.4% | — | 35.8% | 72.1% | — | 84.0% ◇ |
| Qwen2.5-72BAlibaba (Qwen) | 1303 | 49.1% | — | 8.1% | — | — | 5.0% | — | — | 71.6% ◇ |
| Claude Opus 4.7Anthropic | 1501 | 90.2% | 36.2% | 97.8% | 83.5% | 80.2% | — | — | — | — |
| GPT-5.4OpenAI | 1475 | 93.3% | 36.2% | 97.8% | 76.9% | 81.8% | — | — | — | 87.5% ◇ |
| Gemini 3.1 ProGoogle | 1487 | 94.4% | 46.4% | 95.6% | 75.6% | 80.2% | — | — | — | 91.2% |
| Claude Sonnet 4.6Anthropic | 1472 | 87.4% | — | 85.8% | 75.2% | 53.4% | 72.1% | — | — | — |
| Kimi K2.5Moonshot AI | 1450 | 87.6% | 24.4% | 92.2% | 73.8% | 43.2% | 63.3% | — | — | — |
| GPT-5.1OpenAI | 1456 | 87.6% | 23.7% | 88.6% | 68.0% | 47.6% | — | — | — | 86.4% ◇ |
| GPT-5 miniOpenAI | 1390 | 75.0% | 19.4% | 86.7% | 64.7% | 34.8% | — | — | 55.5% | — |
| Claude Opus 4Anthropic | 1426 | 76.3% | 10.7% | 64.4% | 70.7% | — | — | 81.4% ◇ | — | — |
| Gemini 1.5 Pro (Sept 2024)Google | 1351 | 57.2% | 4.6% | 23.1% | — | — | — | — | — | 70.3% |
| Claude 3.5 SonnetAnthropic | 1374 | 54.0% | — | 6.5% | 63.4% | — | — | — | — | 77.6% |
| GPT-4o (May 2024)OpenAI | 1346 | 48.9% | — | 6.3% | 38.8% | — | — | — | — | 72.5% |
| Claude 3 HaikuAnthropic | 1262 | 36.3% | — | 1.8% | 40.6% | — | — | — | — | 42.3% |
| GPT-4 Turbo (Nov 2023)OpenAI | 1325 | 42.4% | — | — | 22.4% | — | — | — | — | 63.7% |
| GLM-5Zhipu AI (Z.ai) | 1458 | 87.8% | — | 80.0% | 72.8% | 52.4% | — | 73.7% | — | 86.0% ◇ |
| Claude Opus 4.6Anthropic | 1504 | 90.5% | 34.4% | 94.4% | 78.7% | 79.8% | — | — | — | — |
| Gemini 3 FlashGoogle | 1473 | 89.4% | — | 95.6% | 75.8% | 64.3% | — | 76.8% | — | — |
| DeepSeek-V3.2DeepSeek | 1425 | 83.4% | — | 87.8% | 70.0% | 39.5% | — | 81.1% ◇ | — | 85.0% ◇ |
| Claude Haiku 4.5Anthropic | 1414 | 71.2% | — | 66.7% | 66.6% | 35.5% | — | — | 68.7% | — |
| GPT-5 nanoOpenAI | 1337 | 69.4% | — | 81.1% | 34.8% | 21.8% | — | — | 51.5% | — |
| Claude Opus 4.1Anthropic | 1450 | 76.8% | 11.5% | 68.9% | 73.3% | 38.0% | — | 82.4% ◇ | — | — |
| GPT-4.1 miniOpenAI | 1383 | 65.8% | — | 44.7% | 23.9% | — | — | 61.4% | 50.5% | — |
| Phi-4Microsoft Research | 1256 | 56.1% | — | 13.8% | — | — | — | — | 28.8% | 70.4% ◇ |
| Llama 3.3 70BMeta | 1318 | 47.4% | — | 5.1% | — | — | — | — | 31.9% | 65.9% |
| Gemini 1.5 Flash (Sep 2024)Google | 1309 | 47.3% | — | 16.3% | — | — | — | — | — | 64.1% |
| Llama 3.1-8BMeta | 1211 | 27.0% | — | 1.7% | — | — | — | — | 25.8% | 36.6% |
| GPT-4o miniOpenAI | 1318 | 37.7% | — | 6.9% | — | — | — | — | — | 63.1% |
| Claude 3 OpusAnthropic | 1322 | 47.2% | — | 4.7% | 15.8% | — | — | — | — | 68.5% |
| Gemini 3.8 FlashGoogle | 1497 | 95.4% | 44.5% | 98.9% | — | — | — | — | — | — |
All benchmarks
Coding
LMArena WebDevActive Arena where two models each build a web app from the same prompt and users vote on the result. 127 models · Best: 1813 LiveCodeBenchNot updated recently Coding problems collected continuously from LeetCode, AtCoder and Codeforces with their release dates, so models can be scored only on problems published after their training cut-off. 22 models · Best: 75.8% HumanEval / HumanEval+Saturated 164 hand-written Python function problems from OpenAI (2021). EvalPlus adds many more tests per problem (HumanEval+) to catch wrong solutions that pass the original tests. 122 models · Best: 89.0%
General
Open LLM Leaderboard v2Archived Hugging Face's leaderboard of open-weight models, scored on IFEval, BBH, MATH Lvl 5, GPQA, MuSR and MMLU-Pro with normalised scores. The leaderboard was archived in 2025 and no longer accepts submissions. 294 models · Best: 48.0% Artificial Analysis Intelligence IndexLink only Independent composite index that runs a battery of evaluations plus price and speed measurements across API models. Link only
Knowledge
MMLUSaturated About 14,000 multiple-choice questions over 57 subjects, from elementary maths to law and medicine. 158 models · Best: 88.1% MMLU-ProActive Harder successor to MMLU: about 12,000 reasoning-heavy questions in 14 disciplines with 10 answer options. 243 models · Best: 91.2% Humanity's Last ExamActive About 2,500 expert-written questions across 100+ subjects from the Center for AI Safety and Scale AI, filtered to stump frontier models at creation time. 46 models · Best: 54.8%
Agentic coding
SWE-bench VerifiedContested 500 human-validated GitHub issues from 12 Python repositories; the system must produce a patch that makes the hidden tests pass. 81 models · Best: 83.5% SWE-bench ProLink only Scale AI's harder, longer-horizon successor to SWE-bench with a 731-task public split and private/commercial splits. Link only Terminal-Bench 2.0Active Hard, realistic tasks an agent must finish in a sandboxed terminal (Stanford × Laude Institute). 45 models · Best: 84.7%
Maths
AIME (OTIS Mock AIME 2024–2025)Active AIME-style competition maths (integer answers 0–999). We use Epoch AI's runs on the OTIS Mock AIME 2024–2025 papers, which are less exposed than past official AIME papers. 200 models · Best: 100.0% MATH Level 5Saturated The hardest (level 5) problems of the MATH competition dataset (Hendrycks et al., 2021). 94 models · Best: 98.1%
Abstract reasoning
ARC-AGI-1Saturated Abstract grid puzzles (François Chollet) that test learning a new rule from a few examples. 6 models · Best: 98.5% ARC-AGI-2Active Harder second version of ARC-AGI (2025), designed so that brute-force search is not enough. 4 models · Best: 95.0% ARC-AGI-3Active Interactive, game-like third version of ARC-AGI that tests exploration and learning inside novel environments. 3 models · Best: 99.9%
Computer use
Tool use & agents
τ²-benchActive Sierra's customer-service agent benchmark: the agent follows a policy and uses tools while an LLM-simulated user talks to it (retail, airline, telecom, banking). 43 models · Best: 85.3% BFCL v4 (function calling)Active Berkeley Function Calling Leaderboard: single, parallel and multi-turn tool calls, web search, memory and relevance detection. 80 models · Best: 77.5%
Chinese
CMMLUNot updated recently Chinese multitask knowledge benchmark: 67 subjects including China-specific topics. 12 models · Best: 90.3% C-EvalNot updated recently Chinese evaluation suite of about 13,900 multiple-choice questions over 52 disciplines and four difficulty levels. 11 models · Best: 68.7% SuperCLUELink only Periodic Chinese general-capability evaluation of domestic and international models, published as reports. Link only
How we collect
- We read official data files and public datasets (Hugging Face, GitHub, CSV downloads) and never invent or estimate scores.
- Model names from different leaderboards are merged into one model page (for example dated versions and reasoning-effort variants).
- Leaderboards that restrict reuse are cited with their top entries only, or linked without copying.
- Data is refreshed by a repeatable script; each run keeps a dated snapshot so changes can be compared over time.
Source type
- Official leaderboard
- The benchmark maintainers' own leaderboard
- Independent run
- An independent evaluator ran the model itself (e.g. Epoch AI)
- Vendor-reported
- Self-reported or submitted by the model developer
- Paper / tech report
- Reported in a paper or technical report (usually by the developer)
- Third-party compilation
- Compiled by Epoch AI; the original source is not given per row
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| LMArena leaderboard dataset LMArena, Arena Leaderboard Dataset (Hugging Face), CC BY 4.0 | CC BY 4.0 | Open licence – full data with attribution | Oct 10, 2026 ⚠︎ |
| Epoch AI Benchmarking Hub Epoch AI, 'Capabilities & benchmarking', https://epoch.ai/benchmarks (CC BY 4.0) | CC BY 4.0 | Open licence – full data with attribution | Oct 22, 2026 |
| SWE-bench leaderboards (swebench.com) SWE-bench team, swebench.com leaderboards | No explicit licence (NOASSERTION) – factual scores only, with citation | No explicit licence – factual scores only, with citation | Oct 22, 2026 |
| LiveCodeBench leaderboard LiveCodeBench (Jain et al.), livecodebench.github.io | No explicit licence on the site repo (benchmark code MIT) – factual scores only, with citation | No explicit licence – factual scores only, with citation | Oct 22, 2026 |
| EvalPlus leaderboard EvalPlus team, evalplus.github.io | Apache-2.0 (site repo) | Open licence – full data with attribution | Oct 22, 2026 |
| Berkeley Function Calling Leaderboard (BFCL) Gorilla / BFCL team, UC Berkeley | Apache-2.0 (gorilla repo) | Open licence – full data with attribution | Oct 22, 2026 |
| τ²-bench leaderboard submissions Sierra Research, τ²-bench | MIT (repo) | Open licence – full data with attribution | Oct 22, 2026 |
| MMLU-Pro leaderboard (TIGER-Lab) TIGER-Lab, MMLU-Pro leaderboard | Apache-2.0 (dataset) | Open licence – full data with attribution | Oct 22, 2026 |
| Hugging Face Open LLM Leaderboard v2 (archived) Hugging Face, Open LLM Leaderboard (archived 2025) | No licence tag – factual scores only, with citation | No explicit licence – factual scores only, with citation 4576 rows in archive; 353 official-developer models kept | Oct 22, 2026 |
| ARC Prize leaderboard ARC Prize Foundation, arcprize.org/leaderboard | ARC Prize Terms: no copying/aggregation for commercial purposes without permission | Restricted – only top entries cited top 10 per set only (terms) | Oct 22, 2026 |
| CMMLU leaderboard (GitHub README) Li et al., CMMLU (CC BY-NC-SA 4.0) | CC BY-NC-SA 4.0 | Restricted – only top entries cited | Oct 22, 2026 |
| C-Eval (GitHub README, initial release table) Huang et al., C-Eval | Code MIT; dataset CC BY-NC-SA 4.0 | Restricted – only top entries cited | Oct 22, 2026 |
| Artificial Analysis Artificial Analysis | Free API requires an API key and attribution; no open bulk data | Not collected – link only | — |
| SuperCLUE CLUE / SuperCLUE | No open data interface | Not collected – link only | — |
| Scale AI SEAL leaderboards (SWE-bench Pro, HLE) Scale AI | No open data interface | Not collected – link only | — |
For AI agents
The same data as JSON: benchmark metadata with licences, every model's best scores and per-benchmark result lists with source links.
/api/benchmarks.json/api/benchmarks/matrix.json/api/benchmarks/<id>.json/llms.txt
FAQ
Which AI model is best right now?
It depends on the task. Human-preference arenas (LMArena), science questions (GPQA Diamond), agentic coding (SWE-bench, Terminal-Bench), computer use (OSWorld) and tool calling (BFCL, τ²-bench) often rank models differently. Use the matrix to filter by benchmark type and compare models side by side.
Why do vendor-reported scores differ from leaderboard scores?
Developers often test with their own prompts, scaffolds, tool access or reasoning budgets, and pick the settings they report. Official leaderboards and independent evaluators such as Epoch AI use one fixed setup for all models. We label each number's source type so you can tell them apart.
Is SWE-bench Verified still a good benchmark?
Use it with care. In February 2026 OpenAI stopped reporting SWE-bench Verified, citing contamination and flawed tests in an audit of hard tasks. It later also cautioned that about 30% of SWE-bench Pro tasks look broken. Scores also depend heavily on the agent scaffold.
Is the Hugging Face Open LLM Leaderboard still updated?
No. It was archived in 2025 and no longer accepts submissions, so it has no recent models. We mark it as archived.
How often is this data updated?
A collection script re-reads every source and stores a dated snapshot. The date at the top of each page shows the latest run; each source also shows when it was last fetched.