Benchmarks · read 20 November 2026

Benchmarks

What public leaderboards publish about the models we list, from the ones whose terms let us republish it, with who ran every number. 18 benchmarks, 677 results on 94 models across 83 listings: 637 run by someone with nothing to sell in the model, 40 submitted by the model's maker.

Every benchmark

BenchmarkAreaUnitRun byModelsBestNewest resultStatus
Berkeley Function Calling Leaderboard (BFCL)Agent tasks% overall accuracyUC Berkeley (Gorilla)5Qwen3-30B-A3B-Instruct-2507 41.4%2025-12-16Stale
METR time horizonsAgent tasksminutes at 50% successMETR1Gemini 3.1 Pro (preview) 384 min2026-09-08Discontinued
Terminal-BenchAgent tasks% of trials passedStanford and the Laude Institute (the Terminal-Bench maintainers)33Claude Opus 5.5 64.8%2026-11-20Current
τ²-benchAgent tasks% of tasks passed on the first try (pass^1)Sierra6Kimi K3 37.1%2026-10-29Current, 4 submitted
Aider PolyglotCoding% of exercises solved in two triesAider (Paul Gauthier)1gpt-oss-120b 41.8%2025-10-03Stale
MirrorCodeCoding% of tasks solvedEpoch AI19Claude Sonnet 5.5 97.8%2026-11-17Current, near its ceiling
SWE-bench VerifiedCoding% of issues resolvedEpoch AI1DeepSeek V4 Pro 77.6%2026-06-25Current
FrontierMath (Tiers 1-3)Reasoning and knowledge% of problems solvedEpoch AI36Claude Fable 5.2 95.1%2026-11-19Current, near its ceiling
FrontierMath Tier 4Reasoning and knowledge% of problems solvedEpoch AI32Claude Fable 5.2 100.0%2026-11-19Current, near its ceiling
GPQA DiamondReasoning and knowledge% of questions correctEpoch AI45GPT-6.2 Astra 96.8%2026-11-19Current, near its ceiling
OTIS Mock AIME 2024-2025Reasoning and knowledge% of problems solvedEpoch AI49Gemini 4 Neon 100.0%2026-11-19Current, near its ceiling
SimpleQA VerifiedReasoning and knowledge% of questions correctEpoch AI30Claude Fable 5.2 94.0%2026-11-19Current
MTEB English v2Embeddingsmean task score (0 to 100)MTEB maintainers (an open community project started at Hugging Face), from results submitted by pull request11Qwen3-Embedding-8B 75.22026-11-20Current, 16 submitted
MTEB English v2 retrievalEmbeddingsmean nDCG@10 (0 to 100)MTEB maintainers (an open community project started at Hugging Face), from results submitted by pull request12Qwen3-Embedding-8B 69.42026-11-20Current, 20 submitted
LMArena Agent ArenaArenas: people's votesIPS scoreLMArena40Claude Fable 5.2 0.2742026-11-18Current
LMArena Code Arena (WebDev)Arenas: people's votesarena rating (Bradley-Terry)LMArena53Claude Fable 5.2 18232026-11-19Current
LMArena Text ArenaArenas: people's votesarena rating (Bradley-Terry)LMArena64Gemini 4 Carbon 15272026-11-19Current
LMArena Vision ArenaArenas: people's votesarena rating (Bradley-Terry)LMArena29Claude Fable 5.2 13052026-11-19Current

Models counts the models we list with a result. Best is the top result among them, on the version of the test most of them took. Stale means nothing new in six months; near its ceiling, that the best result is within five points of a perfect score, so the test no longer tells the top models apart.

Every model against every benchmark

Each model's best result on the version of each test most models took, so a model that only took another version shows nothing here; its benchmark's page has every version. Shaded results were submitted by the model's maker; the rest were run independently.

ModelBerkeley Function Calling Leaderboard (BFCL)METR time horizonsTerminal-Benchτ²-benchAider PolyglotMirrorCodeSWE-bench VerifiedFrontierMath (Tiers 1-3)FrontierMath Tier 4GPQA DiamondOTIS Mock AIME 2024-2025SimpleQA VerifiedMTEB English v2MTEB English v2 retrievalLMArena Agent ArenaLMArena Code Arena (WebDev)LMArena Text ArenaLMArena Vision Arena
Gemini 3.1 Pro (preview)384 min2.7%26.0%8.9%59.6%26.8%94.4%95.6%73.5%-0.057144614871280
Claude Opus 5.564.8%95.6%91.2%95.0%90.6%100.0%72.2%0.186181815061293
Claude Sonnet 5.561.8%97.8%88.8%80.5%95.6%100.0%46.5%0.156177114731264
DeepSeek V4.1 Flash40.0%64.4%71.2%46.3%93.2%99.4%42.9%0.008165914751262
GPT-6 Luna16.4%26.7%79.0%56.1%90.5%98.9%41.4%0.006157914431234
GPT-6 Sol49.4%48.9%89.8%90.0%94.3%100.0%60.7%0.065167414581258
GPT-6.1 Sol58.2%61.1%93.7%100.0%95.4%100.0%73.9%0.113174414831290
GPT-6.1 Terra51.8%76.7%90.9%92.7%94.3%100.0%57.3%0.105168414741263
Gemini 3.8 Flash19.1%38.9%68.4%21.9%95.4%98.9%69.7%0.015158414971293
Kimi K322.4%37.1%54.4%72.2%39.0%93.1%97.2%50.6%-0.00116581488
Claude Fable 5.157.9%93.3%90.2%87.8%100.0%70.8%0.137174115031290
Claude Fable 5.295.1%100.0%93.9%99.7%94.0%0.274182315031305
Claude Haiku 5.540.0%78.9%75.1%46.3%89.7%98.9%23.8%-0.00615961421
GPT-6 Astra60.0%93.7%97.6%95.8%100.0%75.6%0.146176314751281
GPT-6.2 Astra67.8%92.6%100.0%96.8%100.0%78.5%0.15217591468
Gemini 4 Argon54.9%58.9%86.0%78.0%96.4%100.0%76.0%0.10316811520
Grok 4.754.4%53.0%17.1%92.7%98.1%56.0%0.015163814431250
Inkling0.0%25.0%33.3%4.9%88.3%88.9%40.3%-0.18514141442
Inkling-Small0.9%46.3%17.1%88.5%90.0%19.1%-0.185140814061208
Kimi K3.1 Pro60.9%43.3%78.2%65.8%94.2%100.0%62.2%0.03916831482
DeepSeek V4 Pro6.7%77.6%45.3%2.4%90.9%96.7%47.0%15811464
GLM-5.341.8%68.8%29.3%90.9%91.1%41.0%-0.00016221482
GLM-5.3-Flash36.7%55.8%17.1%90.2%93.9%-0.039161014751274
Gemini 4 Carbon90.5%95.1%96.5%100.0%77.7%0.16117691527
Gemini 4 Hydrogen88.1%87.8%96.8%100.0%70.9%0.13517011490
Gemini 4 Neon91.2%97.6%96.3%100.0%75.2%0.18417631513
Muse Spark 1.314.6%54.4%74.4%46.3%99.2%0.011165414981288
Grok 4.342.8%14.6%88.8%93.3%33.2%135714421241
Mistral Large 427.9%90.2%99.4%22.4%-0.071157014231233
Muse Wonder47.0%45.6%82.1%70.7%99.4%0.00616481511
Gemini 3.5 Flash-Lite26.0%0.0%83.3%71.1%143214551266
MiniMax M37.3%90.9%71.1%-0.109148414341231
Qwen3.5-397B-A17B9.8%31.2%86.4%88.9%140014421245
Muse Spark Open69.8%31.7%88.6%98.1%56.6%1527
DeepSeek-V4-Flash-073157.5%24.4%91.0%94.4%33.6%
Gemma 4 26B A4B IT73.2%82.2%135814381241
Nemotron 3.5 Super VL0.3%-0.173140214161211
GPT-6.2 Sol0.11317541465
Mistral Medium 3.5-0.145126314261199
Qwen3-30B-A3B-Instruct-250741.4%55.6%62.2%1382
Qwen3.8-27B-0.055158914351244
Step 5 Preview-0.051157014491246
gpt-oss-120b41.8%75.8%88.9%1352
DeepSeek-V3.213621425
GLM-4.7-Flash60.5%58.3%1365
Gemini 3.1 Flash-Lite27.7%81.8%80.0%
Gemma 4 31B137014521265
Gemma 4 31B IT75.8%73.3%10.4%
Hy3-0.08615061454
Inkling 2 Small0.01516441412
Llama 3.3 70B31.9%47.4%5.1%
Mistral Large 3123314151200
Muse Ember0.0101609
Nemotron 3 Ultra 550B A55B85.3%86.7%1427
Qwen3-4B-Instruct-250735.7%45.8%52.2%
Qwen3.5-122B-A10B136014161226
Qwen3.6-27B35.1%85.9%91.1%
Qwen3.6-35B-A3B20.4%84.8%86.7%
Tencent Hy439.1%0.0611650
gpt-oss-20b60.8%65.3%1318
Amazon Nova 2 Lite27.1%1335
Cohere Embed and Rerank
EmbeddingGemma 2
EmbeddingGemma 300M
Gemini Embedding
Granite Embedding Small English R2
Jina Embeddings and Reranker
MiniMax-M2.513861391
MiniMax-M2.713981415
MiniMax-M3.1-Flash-0.0371582
Nomic Embed
Qwen3-Embedding-0.6B
Qwen3-Embedding-4B
Qwen3-Embedding-8B
Qwen3.5-9B79.0%61.7%
Qwen3.8-Flash-Next-0.0491626
jina-embeddings-v5-omni-small
Beam 1.06.4%
Command R1251
Command R+ 08-20241276
Command R7B32.1%
DeepSeek V3.11417
GLM-4.5-Air1373
GPT-Live
Granite 4.2 30B1340
Granite 4.2 3B1289
Granite 4.2 8B1290
Laguna M.11350
Nemotron 3 Nano 30B A3B1314
Nemotron 3 Super 120B A12B1360
Nemotron 3.5 Lightning 30B A3B1352
OpenAI embeddings
Qwen3.5-4B55.8%
Step 3.5 Flash1393

Who ran it

  • Independent. A party with nothing to sell in the model ran it and published how: Epoch AI, LMArena's votes, the Terminal-Bench maintainers, METR.
  • Submitted by the maker. The model's maker ran it and the leaderboard took the number, as MTEB and some of τ²-bench's rows are. Shown, and marked.
  • The maker's claim. A number only on the maker's own pages, a launch post or a model card. None is collected yet.
  • Measured by Anchor Terminal. Our own runs, such as the lab's evaluation. Not in these files yet.

A rank on these pages is among the models we list with a result, at each model's best setting, not a place on the whole leaderboard. A model offered by several hosts is one set of results, shown on each of its listings. These results sit beside our grades and don't move them: how we grade a listing is a separate method, at /benchmark/.

Not here yet

SWE-bench's own leaderboard, Humanity's Last Exam and ARC-AGI, until their makers allow republishing (their terms don't today); Epoch AI's own SWE-bench Verified runs stand in for the first. Artificial Analysis, the speech and text-to-speech arenas, the benchmarks for coding agents, browsers, search and memory, and the numbers vendors publish about themselves come next.

Credits

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.