LLM Benchmarks

Source-backed measurements

Compare 100 AI models across overall capability, knowledge, coding and agents—with sources, test dates and transparent comparison rules.

Page updated
EN
♡ Support
New

First-of-its-kind Local GPU Benchmarks: We test compact open-weight models (4B–35B) on our own hardware (2× RTX 5060 Ti, 32 GB VRAM), with and without thinking. Test scope, available runtime, speed and time efficiency are shown.

Explore local GPU benchmarks →
Benchmarks
Models
Filters
Models
Cell view
Data quality
Compare models
Available models: 100
100 of 100 models · 18 benchmarks
Read the table
Find a model · Click a header to sort · Click a score for its source Line: green newest · blue current · ochre after 6 months≡ Shared rank: the lead is smaller than a single solved task#3/13 Window place: ranked only inside its own task windowDate Date label = date basisS·A+ Source authority: A+ highest · A strong · B solid · C limitedT·Max Highest measured thinking16 18 columns of benchmarks and derived indices. Categories show all their tests.Ranks and shades use all rankable models in each benchmark, including hidden models. Five bands are used from ten results; smaller cohorts remain neutral. These ranks are derived, not source measurements.ORG Benchmark organizer · PROV Model provider · 3P Official cross-provider comparison credible sources disagree for the same disclosed protocol directly comparable under the same protocol, with possible benchmark contamination different protocolMeasured quality targetsComparable cells: 90.5% · target ≥85%Fresh essential cells: 97.2% · target ≥90%Essential columns with ≥60 models: 5/16
Share & export

For each test setup, we show the highest disclosed thinking level. It may not produce the highest score. Levels are not equal compute budgets across providers.

Latest verified refresh No material score or table changeOpen machine-readable change feed
Cell view
Top 10%Top 25%MiddleBottom quarterBottomRanks and colors refer to all models per benchmark. Model filters only hide rows.Highest measured thinking

Current, source-attributed results for 100 language models. Columns are sortable. Comparable values are ranked; SWE-Pro can also rank clearly labelled official system reports, with differing protocols marked ≈.
OverviewKnowledge and reasoningCodingAgents and tools
⁨ECI⁩ rank ↓ · ModelECI, Overview, Epoch Capabilities Index: 86/100 measured models; 86 comparable, 1 sharing a place. Canonical protocol: Epoch ECI snapshot 2026-11-04 · official-composite-fit · Epoch Capabilities Index. Best comparable first.InfoAA Index, Overview, Artificial Analysis Intelligence Index: 28/100 measured models; 28 comparable, 1 sharing a place. Canonical protocol: Artificial Analysis Intelligence Index v5.0 · intelligence-index-v5-0 · Artificial Analysis · Intelligence Index evaluations. Sort.InfoCoding Index, Overview, LLM Benchmarks Coding Index: 81/100 measured models; 81 comparable, 6 sharing a place. Canonical protocol: anchor-v2 · ≥ 2 tests · LLM Benchmarks. Sort.InfoAgent Index, Overview, LLM Benchmarks Agent Index: 88/100 measured models; 88 comparable, 2 sharing a place. Canonical protocol: anchor-v2 · ≥ 2 tests · LLM Benchmarks. Sort.InfoArena, Overview, Arena Text: 80/100 measured models; 80 comparable, 25 sharing a place. Canonical protocol: 2026-10-28 · overall-style-control · Arena Text. Sort.InfoLiveBench, Overview, LiveBench: 57/100 measured models; 57 comparable. Canonical protocol: 2026-06-25 · overall-seven-category · LiveBench. Sort.InfoHLE Full, Knowledge and reasoning, Humanity's Last Exam · Full multimodal: 54/100 measured models; 16 comparable, 38 with a different protocol, 1 sharing a place. Canonical protocol: HLE finalized 2,500 · Scale protocol 2025-04-03 · full-multimodal-2500-no-tools · Scale AI Labs · Humanity's Last Exam Full. Sort.InfoGPQA, Knowledge and reasoning, GPQA Diamond: 92/100 measured models; 53 comparable, 39 with a different protocol, 26 sharing a place. Canonical protocol: GPQA Diamond v1 · Vals dual-prompt · diamond-198-zero-and-five-shot-cot · Vals AI GPQA Diamond. Sort.InfoMMLU-Pro, Knowledge and reasoning, MMLU-Pro: 63/100 measured models; 54 comparable, 9 with a different protocol, 13 sharing a place. Canonical protocol: Vals MMLU-Pro v1 · vals-mmlu-pro-overall · Vals AI · MMLU-Pro. Sort.InfoFrontierMath T1–3, Knowledge and reasoning, FrontierMath T1–3: 68/100 measured models; 68 comparable, 19 sharing a place. Canonical protocol: 2.0.0 · private-285-of-295-problem-tiers-1-3-set · Epoch AI · FrontierMath Tiers 1–3 v2. Sort.InfoSWE-bench, Coding, SWE-bench Verified: 55/100 measured models; 55 comparable, 18 sharing a place. Canonical protocol: SWE-bench Verified v1 · verified-500-vals-mini-swe-agent · Vals AI SWE-bench Verified. Shows only Vals AI's uniform 500-task mini-swe-agent/Bash evaluation. This is the default directly comparable view. Sort.InfoSWE-rebench, Coding, SWE-rebench v2: 35/100 measured models; 13 comparable, 8 directly comparable with a contamination warning, 22 with a different protocol. Canonical protocol: rolling-v2 · 2026-05-15/2026-07-01 · SWE-rebench. Shows the latest shared SWE-rebench task window. Contamination warnings remain visible and do not change the task-window match. 2026-05-15/2026-07-01 Sort.InfoLiveCodeBench, Coding, LiveCodeBench: 60/100 measured models; 52 comparable, 8 with a different protocol, 6 sharing a place. Canonical protocol: Vals LiveCodeBench v1 · vals-livecodebench-overall · Vals AI · LiveCodeBench. Sort.InfoSciCode, Coding, SciCode: 84/100 measured models; 77 comparable, 7 with a different protocol, 10 sharing a place. Canonical protocol: SciCode current leaderboard · scientific-research-coding-subproblems · Epoch AI mirror · SciCode. Sort.InfoTerminal 2.1, Agents and tools, Terminal-Bench 2.1: 69/100 measured models; 57 comparable, 12 with a different protocol, 13 sharing a place. Canonical protocol: Vals Terminal-Bench 2.1 v2.1 · vals-terminal-bench-2-1-overall · Vals AI · Terminal-Bench 2.1. Sort.InfoMCP-Atlas, Agents and tools, MCP-Atlas · Scale: 24/100 measured models; 23 comparable, 1 with a different protocol, 2 sharing a place. Canonical protocol: MCP-Atlas April 2026 rescored protocol · all-1000-public-500-private-500 · Scale AI Labs · MCP-Atlas. Sort.InfoAPEX, Agents and tools, APEX-Agents: 51/100 measured models; 51 comparable, 9 sharing a place. Canonical protocol: APEX-Agents · Epoch evaluation since 2026-09-30 · 480-professional-tasks-33-worlds · Epoch AI mirror · APEX-Agents. Sort.InfoGDPval-AA v2.1, Agents and tools, GDPval-AA v2.1: 79/100 measured models; 79 comparable, 2 sharing a place. Canonical protocol: GDPval-AA v2.1 (Crowd-BT Elo, DeepSeek V4.1 Flash (max) at 1600) · 220-tasks-elo-v2-1 · Artificial Analysis · Intelligence Index evaluations. Sort.Info
Rank 1GPT 6 astraOpenAI13/18 measured · 12 comparable
Rank 2Claude Opus 5.5Anthropic12/18 measured · 11 comparable
Rank 3GPT 6.1 TerraOpenAI9/18 measured · 8 comparable
Rank 4GPT 6.1 SolOpenAI11/18 measured · 10 comparable
Rank 5Claude Sonnet 5.5Anthropic13/18 measured · 12 comparable
Rank 6Claude Fable 5.1Anthropic16/18 measured · 16 comparable
Rank 7Gemini 4 ArgonGoogle11/18 measured · 10 comparable
Rank 8Claude Opus 5Anthropic17/18 measured · 16 comparable
Rank 9GPT 6 SolOpenAI11/18 measured · 10 comparable
Rank 10GPT 5.5 ProOpenAI3/18 measured · 2 comparable
Rank 11Claude Fable 5Anthropic16/18 measured · 16 comparable
Rank 12GPT 5.6 SolOpenAI16/18 measured · 16 comparable
Rank 13Muse wonderMeta7/18 measured · 6 comparable
Rank 14Claude haiku 5.5Anthropic12/18 measured · 11 comparable
Rank 15GPT 5.6 TerraOpenAI14/18 measured · 14 comparable
Rank 16GPT 5.4 ProOpenAI4/18 measured · 3 comparable
Rank 17Claude Opus 4.8Anthropic17/18 measured · 15 comparable
Rank 18GPT 5.5OpenAI17/18 measured · 15 comparable
Rank 19DeepSeek V4.1 Flash (2026-10-17)DeepSeek10/18 measured · 8 comparable
Rank 20Kimi K3Moonshot AI16/18 measured · 15 comparable
Rank 21Gemini 3.7 FlashGoogle14/18 measured · 14 comparable
Rank 22Qwen 3.8 MaxAlibaba14/18 measured · 14 comparable
Rank 23Gemini 3.8 FlashGoogle16/18 measured · 16 comparable
Rank 24Muse Spark 1.3Meta11/18 measured · 10 comparable
Rank 25GPT 5.4OpenAI16/18 measured · 15 comparable
Rank 26Claude Sonnet 5Anthropic16/18 measured · 15 comparable
Rank 27Grok 4.6xAI14/18 measured · 14 comparable
Rank 28GPT 6 LunaOpenAI12/18 measured · 11 comparable
Rank 29Claude Opus 4.7Anthropic17/18 measured · 16 comparable
Rank 30GPT 5.6 LunaOpenAI14/18 measured · 14 comparable
Rank 31GPT 5.2 ProOpenAI2/18 measured · 2 comparable
Rank 32Qwen 3.8 Max (0902)Alibaba8/18 measured · 6 comparable
Rank 33GLM 5.3Z.ai17/18 measured · 16 comparable
Rank 34DeepSeek V4.1 FlashDeepSeek11/18 measured · 9 comparable
Rank 35DeepSeek V4 Pro (2026-08-13)DeepSeek16/18 measured · 15 comparable
Rank 36Claude Opus 4.6Anthropic14/18 measured · 13 comparable
Rank 37Gemini 3.1 ProGoogle17/18 measured · 16 comparable
Rank 38Muse Spark 1.2Meta12/18 measured · 11 comparable
Rank 39Muse Spark openMeta8/18 measured · 5 comparable
Rank 40Gemini 3.5 FlashGoogle17/18 measured · 15 comparable
Rank 41DeepSeek V4 Flash (2026-07-31)DeepSeek13/18 measured · 12 comparable
Rank 42Gemini 3.6 FlashGoogle14/18 measured · 14 comparable
Rank 43Grok 4.7xAI12/18 measured · 11 comparable
Rank 44Grok 4.5xAI15/18 measured · 15 comparable
Rank 45Qwen 3.7 MaxAlibaba13/18 measured · 13 comparable
Rank 46Muse Spark 1.1Meta15/18 measured · 14 comparable
Rank 47GPT 5.2OpenAI12/18 measured · 11 comparable
Rank 48Gemini 3 ProGoogle11/18 measured · 10 comparable
Rank 48GLM 5.3 FlashZ.ai16/18 measured · 15 comparable
Rank 50Claude Sonnet 4.6Anthropic16/18 measured · 14 comparable
Rank 51Grok 4.20xAI10/18 measured · 10 comparable
Rank 52Muse SparkMeta11/18 measured · 11 comparable
Rank 53GLM 5.2Z.ai744B · 40B active17/18 measured · 16 comparable
Rank 54Gemini 3 FlashGoogle12/18 measured · 10 comparable
Rank 55Qwen 3.8 27BAlibaba27B15/18 measured · 14 comparable
Rank 56Kimi K2.6Moonshot AI1T · 32B active15/18 measured · 13 comparable
Rank 57Mistral Large 4Mistral AI9/18 measured · 8 comparable
Rank 58Inkling SmallThinking Machines Lab276B · 12B active15/18 measured · 14 comparable
Rank 59Kimi K2.7 CodeMoonshot AI1T · 32B active12/18 measured · 11 comparable
Rank 60Claude Opus 4.5Anthropic12/18 measured · 11 comparable
Rank 61GPT 5.1OpenAI11/18 measured · 11 comparable
Rank 62GLM 5.1Z.ai744B · 40B active16/18 measured · 14 comparable
Rank 63Qwen 3.6 Max PreviewAlibaba4/18 measured · 3 comparable
Rank 64DeepSeek V4 Pro (2026-04-22)DeepSeek1.6T · 49B active15/18 measured · 14 comparable
Rank 65Grok 4.3xAI14/18 measured · 13 comparable
Rank 66GPT 5.4 MiniOpenAI13/18 measured · 13 comparable
Rank 67InklingThinking Machines Lab975B · 41B active17/18 measured · 16 comparable
Rank 68Kimi K2.5Moonshot AI1T · 32B active13/18 measured · 12 comparable
Rank 69Qwen 3.6 PlusAlibaba13/18 measured · 13 comparable
Rank 70Qwen 3.7 PlusAlibaba10/18 measured · 8 comparable
Rank 71MiniMax M3MiniMax16/18 measured · 15 comparable
Rank 72Qwen 3.5 397B-A17BAlibaba397B · 17B active12/18 measured · 6 comparable
Rank 73Claude Sonnet 4.5Anthropic13/18 measured · 12 comparable
Rank 74Qwen 3.6 27BAlibaba27B14/18 measured · 9 comparable
Rank 75DeepSeek V4 Flash (2026-04-22)DeepSeek284B · 13B active9/18 measured · 5 comparable
Rank 76GPT 5.4 NanoOpenAI13/18 measured · 13 comparable
Rank 77GLM 5Z.ai9/18 measured · 7 comparable
Rank 78Gemini 3.5 Flash LiteGoogle15/18 measured · 15 comparable
Rank 79Claude Opus 4.1Anthropic7/18 measured · 6 comparable
Rank 80Qwen 3.6 35B-A3BAlibaba35B · 3B active12/18 measured · 7 comparable
Rank 81Gemma 4 31B ITGoogle30.7B7/18 measured · 3 comparable
Rank 82Qwen 3.5 35B-A3BAlibaba35B · 3B active12/18 measured · 7 comparable
Rank 83GPT 5.5 InstantOpenAI5/18 measured · 4 comparable
Rank 84Qwen 3.5 9BAlibaba9B8/18 measured · 4 comparable
Rank 85Qwen 3 32BAlibaba32.8B6/18 measured · 5 comparable
Rank 86Qwen 3 14BAlibaba14.8B5/18 measured · 4 comparable
14 models without a directly comparable ECI score
No rankGemini 4 carbonGoogle6/18 measured · 6 comparable
No rankMiMo V2.6 ProXiaomi8/18 measured · 8 comparable
No rankMuse wonder 1.0Meta6/18 measured · 6 comparable
No rankMiMo V2.5 ProXiaomi12/18 measured · 11 comparable
No rankstep 5 PreviewStepFun6/18 measured · 6 comparable
No rankMiMo V2.6 FlashXiaomi7/18 measured · 7 comparable
No rankHY3Tencent7/18 measured · 3 comparable
No rankGrok 4.1xAI1/18 measured · 1 comparable
No rankQwen 3.5 27BAlibaba27B7/18 measured · 2 comparable
No rankQwen 3.8 Flash NextAlibaba125B · 6B active7/18 measured · 3 comparable
No rankGemma 4 12B ITGoogle12B4/18 measured · 0 comparable
No rankGemma 4 26B-A4B ITGoogle25.2B · 3.8B active4/18 measured · 0 comparable
No rankQwen 3.8 2.4T-A95BAlibaba2.4T · 95B active7/18 measured · 3 comparable
No rankGemma 4 31BGoogle7/18 measured · 3 comparable