Of the 901 benchmarks labs named in 2026, 585 were named by just one lab
We counted every benchmark named in the 266 launch posts labs published in 2026, and in the 139 system cards, model cards and technical reports published with them. Each shape below is one benchmark. The more labs named it, the further right it sits, and the more releases named it, the bigger it is.
The frontier labs are the nine whose best 2026 models score highest on Epoch AI's Capabilities Index, as of November 2026: Anthropic, OpenAI, Google DeepMind, Moonshot AI, Meta, DeepSeek, xAI, Qwen and Z.ai. A frontier model ranked in the top ten on Epoch AI's Capabilities Index among the models out the day it was released.
The data for this view didn't load, so the page shows 2026 for every lab. Try the switches again in a moment.
- Knowledge and language
- Math and science
- Coding
- Agents and tools
- Professional work
- Vision, audio and robotics
- Safety and security
- Other
releases named it
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 34% of the benchmarks labs named had come out in the year before the release. In the second half of 2026 so far, 63% had. The ten most-named benchmarks went from 35% of mentions to 20%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 39% of the benchmarks labs named had come out in the year before the release. In the second half of 2026 so far, 65% had. The ten most-named benchmarks went from 30% of mentions to 14%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 40% of the benchmarks labs named for frontier models had come out in the year before the release. In the second half of 2026 so far, 77% had. The ten most-named benchmarks went from 32% of mentions to 25%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 46% of the benchmarks labs named for frontier models had come out in the year before the release. In the second half of 2026 so far, 76% had. The ten most-named benchmarks went from 28% of mentions to 11%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 37% of the benchmarks frontier labs named had come out in the year before the release. In the second half of 2026 so far, 73% had. The ten most-named benchmarks went from 34% of mentions to 23%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 41% of the benchmarks frontier labs named had come out in the year before the release. In the second half of 2026 so far, 72% had. The ten most-named benchmarks went from 29% of mentions to 13%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 42% of the benchmarks frontier labs named for frontier models had come out in the year before the release. In the second half of 2026 so far, 77% had. The ten most-named benchmarks went from 31% of mentions to 25%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
How has reporting changed?
In the first half of 2024, 48% of the benchmarks frontier labs named for frontier models had come out in the year before the release. In the second half of 2026 so far, 76% had. The ten most-named benchmarks went from 27% of mentions to 11%.
Released in the past year
Named benchmarks that came out in the 12 months before the release
The ten most named
Mentions that went to the ten benchmarks named most that half year, with versions of a test as one
The lab’s own benchmarks
Named benchmarks that the releasing lab made
Benchmarks per release
The median release, counting each benchmark once
Source: Labs’ launch posts, read by Benchmark Almanac
Do lab numbers match independent runs?
In 2026 so far, labs' own numbers beat an independent run of the same model on the same benchmark in 56% of the pairs where the two differ, by a median of 0.7 points. In 2024 it was 64%. This counts every lab, whatever the switches above say.
The lab’s number is higher
Of the pairs where the two differ
Median gap, in points
The lab’s number less the independent run’s
Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Which benchmarks stay on top?
We track the ten benchmarks labs cite in the most releases each quarter. Only GPQA Diamond has made the top ten in every quarter since 2024. Terminal-Bench has been first for three quarters running.
Source: Labs’ launch posts, read by Benchmark Almanac
Which benchmarks stay on top?
We track the ten benchmarks labs cite in the most frontier-model releases each quarter. GPQA Diamond and SWE-bench Verified each made the top ten in six of the eight quarters since 2025 began, more than any other benchmark.
Source: Labs’ launch posts, read by Benchmark Almanac
Which benchmarks stay on top?
We track the ten benchmarks frontier labs cite in the most releases each quarter. Humanity's Last Exam made frontier labs' top ten in seven of the eight quarters since 2025 began, more than any other benchmark.
Source: Labs’ launch posts, read by Benchmark Almanac
Which benchmarks stay on top?
We track the ten benchmarks frontier labs cite in the most frontier-model releases each quarter. GPQA Diamond and SWE-bench Verified each made frontier labs' top ten in six of the eight quarters since 2025 began, more than any other benchmark.
Source: Labs’ launch posts, read by Benchmark Almanac
How fast do new benchmarks catch on?
DeepSWE reached ten labs 38 days after the first one cited it, the fastest of the 29 new benchmarks that got there. Half of them took 211 days or less. We count every benchmark released since 2025 that at least five labs have cited.
Source: Labs’ launch posts, read by Benchmark Almanac
How fast do new benchmarks catch on?
Humanity's Last Exam is the only new benchmark that reached eight labs, 379 days after the first one cited it. We count every benchmark released since 2025 that at least three labs have cited for frontier models.
Source: Labs’ launch posts, read by Benchmark Almanac
How fast do new benchmarks catch on?
DeepSWE reached all nine frontier labs 48 days after the first one cited it, the fastest of the five new benchmarks that got there. Half of them took 234 days or less. We count every benchmark released since 2025 that at least five frontier labs have cited.
Source: Labs’ launch posts, read by Benchmark Almanac
How fast do new benchmarks catch on?
Humanity's Last Exam is the only new benchmark that reached all eight frontier labs, 379 days after the first one cited it. We count every benchmark released since 2025 that at least three frontier labs have cited for frontier models.
Source: Labs’ launch posts, read by Benchmark Almanac
Which benchmarks do one or two labs carry?
Of the 332 benchmarks named in five or more releases since 2024, 144 owe at least half of those releases to one lab, and 7 more owe four in five to two labs. 41 of the ones a single lab carries are its own. Qwen carries the most that others made, 50, among them C-Eval and SuperGPQA.
Which benchmarks do one or two labs carry?
Of the 491 benchmarks named in five or more releases since 2024, 292 owe at least half of those releases to one lab, and 8 more owe four in five to two labs. 164 of the ones a single lab carries are its own. Qwen carries the most that others made, 50, among them C-Eval and SuperGPQA.
- C-Eval Qwen 50% 36
- SuperGPQA Qwen 56% 36
- MMLU-Redux Qwen 71% 35
- RealWorldQA Qwen 57% 35
- ERQA Qwen 55% 33
- LVBench Qwen 53% 32
- RefCOCO / RefCOCO+ / RefCOCOg Qwen 50% 28
- INCLUDE Qwen 60% 25
- CountBench Qwen 58% 24
- AndroidWorld Qwen 61% 23
- DynaMath Qwen 52% 23
- MMLU-ProX Qwen 61% 23
- MLVU Qwen 77% 22
- MVBench Qwen 64% 22
- MedXpertQA-MM Qwen 52% 21
- LiveBench Qwen 55% 20
- MMLongBench-Doc Qwen 58% 19
- Global MMLU Lite Google DeepMind 67% 18
- V* Qwen 53% 17
- VITA-Bench Qwen 65% 17
- MRCR Google DeepMind 70% its own 37
- C-Eval Qwen 50% 36
- SuperGPQA Qwen 56% 36
- MMLU-Redux Qwen 71% 35
- RealWorldQA Qwen 57% 35
- ERQA Qwen 55% 33
- LVBench Qwen 53% 32
- OpenAI MRCR OpenAI 50% its own 28
- RefCOCO / RefCOCO+ / RefCOCOg Qwen 50% 28
- INCLUDE Qwen 60% 25
- CountBench Qwen 58% 24
- AndroidWorld Qwen 61% 23
- DynaMath Qwen 52% 23
- MMLU-ProX Qwen 61% 23
- MLVU Qwen 77% 22
- MVBench Qwen 64% 22
- HealthBench Hard OpenAI 76% its own 21
- MedXpertQA-MM Qwen 52% 21
- LiveBench Qwen 55% 20
- TroubleshootingBench OpenAI 100% its own 20
Source: Labs’ launch posts, read by Benchmark Almanac
- BBQ (Bias Benchmark for QA) Anthropic 55% 38
- C-Eval Qwen 50% 36
- SuperGPQA Qwen 56% 36
- MMLU-Redux Qwen 71% 35
- RealWorldQA Qwen 57% 35
- Virology Capabilities Test (VCT) Anthropic 50% 34
- ERQA Qwen 55% 33
- LVBench Qwen 53% 32
- StrongREJECT OpenAI 67% 30
- RefCOCO / RefCOCO+ / RefCOCOg Qwen 50% 28
- INCLUDE Qwen 60% 25
- CountBench Qwen 58% 24
- AndroidWorld Qwen 61% 23
- DynaMath Qwen 52% 23
- MMLU-ProX Qwen 61% 23
- MLVU Qwen 77% 22
- MVBench Qwen 64% 22
- MedXpertQA-MM Qwen 52% 21
- LiveBench Qwen 55% 20
- Cybench Anthropic 58% 19
- BBQ (Bias Benchmark for QA) Anthropic 55% 38
- MRCR Google DeepMind 70% its own 37
- C-Eval Qwen 50% 36
- SuperGPQA Qwen 56% 36
- MMLU-Redux Qwen 71% 35
- RealWorldQA Qwen 57% 35
- Virology Capabilities Test (VCT) Anthropic 50% 34
- ERQA Qwen 55% 33
- LVBench Qwen 53% 32
- HealthBench OpenAI 52% its own 31
- Multimodal Troubleshooting Virology OpenAI 100% its own 30
- StrongREJECT OpenAI 67% 30
- Tacit Knowledge and Troubleshooting OpenAI 100% its own 29
- OpenAI MRCR OpenAI 50% its own 28
- ProtocolQA Open-Ended OpenAI 100% its own 28
- RefCOCO / RefCOCO+ / RefCOCOg Qwen 50% 28
- OpenAI Production Benchmarks and safe-completion evals OpenAI 100% its own 26
- INCLUDE Qwen 60% 25
- CountBench Qwen 58% 24
- Image input safety evaluations OpenAI 100% its own 24
Source: Labs’ launch posts, read by Benchmark Almanac
How long do benchmarks stay in launch posts?
Of the benchmarks labs first named in 2025, 33% appeared in only one release, and 37% were still being named a year later. Of those first named in 2026 before May 21st, 43% have appeared in only one release. This counts every lab, whatever the frontier switches say.
Source: Labs’ launch posts, read by Benchmark Almanac
How long do benchmarks stay in launch posts?
Of the benchmarks labs first named in 2025, 31% appeared in only one release, and 34% were still being named a year later. Of those first named in 2026 before May 21st, 37% have appeared in only one release. This counts every lab, whatever the frontier switches say.
Source: Labs’ launch posts, read by Benchmark Almanac
Professional work rose from 1% of what labs reported in 2025 to 10% in 2026
What each of the 16 labs with the most releases named in 2026, by area. Each grid has 100 shapes, one for every 1% of the benchmarks the lab named across its releases. Labs with similar mixes sit side by side. "Their own" is the share the lab made itself, alone or with others, and "Shared" is the share that at least three other labs also named that year.
- Knowledge and language
- Math and science
- Coding
- Agents and tools
- Professional work
- Vision, audio and robotics
- Safety and security
- Other
- Their own
- 21%
- Shared
- 72%
- Per release
- 8.5
- Their own
- 7%
- Shared
- 70%
- Per release
- 28
- Their own
- 6%
- Shared
- 82%
- Per release
- 10
- Their own
- 17%
- Shared
- 72%
- Per release
- 13.5
- Their own
- 7%
- Shared
- 77%
- Per release
- 8
- Their own
- 2%
- Shared
- 76%
- Per release
- 11
- Their own
- 5%
- Shared
- 66%
- Per release
- 42
- Their own
- 16%
- Shared
- 76%
- Per release
- 33
- Their own
- 2%
- Shared
- 66%
- Per release
- 27
- Their own
- 22%
- Shared
- 58%
- Per release
- 37
- Their own
- 6%
- Shared
- 69%
- Per release
- 6
- Their own
- 11%
- Shared
- 67%
- Per release
- 12.5
- Their own
- 15%
- Shared
- 74%
- Per release
- 17
- Their own
- 40%
- Shared
- 60%
- Per release
- 10
- Their own
- 10%
- Shared
- 45%
- Per release
- 6
- Their own
- 0%
- Shared
- 77%
- Per release
- 10.5
Source: Labs’ launch posts, read by Benchmark Almanac
Terminal-Bench was named by 33 of the 50 labs that published scores in 2026
Look up any benchmark and lab. Each dot counts the lab's releases whose launch post or documents named the benchmark, in the benchmark's area color. The columns are the 12 labs with the most releases that year, and the last one adds up every other lab. Point at a dot to see the releases and the scores they reported.
Showing the 207 benchmarks that at least 3 labs named in 2026, out of 901 in total.
| Benchmark | OpenAI | Anthropic | Z.ai | Liquid | xAI | 38 others | Labs | Releases | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1Terminal-BenchStanford + Laude Institute | 10 | 15 | 10 | 13 | 6 | 5 | 9 | 7 | 1 | 3 | 2 | 2 | 51 | 33/50 | 134 |
| 2SWE-bench VerifiedPrinceton NLP + OpenAI | 1 | 11 | 8 | 5 | 2 | 2 | 2 | 2 | 2 | 1 | 37 | 31/50 | 73 | ||
| 3tau-benchSierra Research | 7 | 12 | 5 | 2 | 8 | 3 | 2 | 2 | 1 | 1 | 1 | 26 | 31/50 | 70 | |
| 4AIMEMathematical Association of America (exam), MathArena (evaluation) | 6 | 5 | 1 | 2 | 4 | 4 | 1 | 1 | 4 | 2 | 2 | 34 | 31/50 | 66 | |
| 5GPQA DiamondNYU (Rein et al.) | 8 | 20 | 12 | 5 | 8 | 4 | 3 | 3 | 4 | 2 | 1 | 42 | 29/50 | 112 | |
| 6SWE-Bench ProScale AI | 4 | 11 | 9 | 11 | 1 | 2 | 3 | 2 | 1 | 2 | 31 | 28/50 | 77 | ||
| 7Humanity's Last ExamCenter for AI Safety + Scale AI | 10 | 15 | 5 | 13 | 6 | 6 | 4 | 5 | 2 | 1 | 37 | 27/50 | 104 | ||
| 8LiveCodeBenchUC Berkeley / MIT / Cornell | 7 | 17 | 7 | 2 | 3 | 2 | 1 | 1 | 25 | 26/50 | 65 | ||||
| 9SWE-bench MultilingualSWE-bench team (Princeton/Stanford) | 9 | 12 | 4 | 1 | 2 | 1 | 1 | 1 | 27 | 25/50 | 58 | ||||
| 10IFBenchAllenAI | 4 | 17 | 9 | 1 | 1 | 7 | 2 | 2 | 25 | 24/50 | 68 | ||||
| 11DeepSWEDatacurve | 4 | 6 | 7 | 5 | 1 | 3 | 6 | 5 | 3 | 1 | 18 | 23/50 | 59 | ||
| 12MMLU-ProTIGER-Lab (Waterloo) | 6 | 16 | 7 | 3 | 2 | 1 | 1 | 27 | 22/50 | 63 | |||||
| 13MMMU-ProMMMU team | 10 | 15 | 9 | 2 | 1 | 2 | 1 | 1 | 1 | 22 | 20/50 | 64 | |||
| 14BrowseCompOpenAI | 1 | 4 | 8 | 8 | 5 | 3 | 2 | 2 | 28 | 20/50 | 61 | ||||
| 15AutomationBenchZapier | 5 | 3 | 9 | 8 | 1 | 2 | 4 | 5 | 2 | 1 | 14 | 20/50 | 54 | ||
| 16MCP AtlasScale AI | 2 | 4 | 4 | 6 | 1 | 3 | 4 | 2 | 1 | 17 | 20/50 | 44 | |||
| 17SciCodeSciCode team | 5 | 3 | 7 | 2 | 2 | 22 | 20/50 | 41 | |||||||
| 18GDPval-AAArtificial Analysis | 6 | 1 | 3 | 12 | 3 | 3 | 6 | 2 | 2 | 20 | 19/50 | 58 | |||
| 19HMMTHarvard-MIT Mathematics Tournament (HMMT) | 14 | 2 | 3 | 2 | 1 | 20 | 19/50 | 42 | |||||||
| 20CharXivPrinceton NLP | 6 | 17 | 1 | 7 | 2 | 1 | 3 | 20 | 18/50 | 57 | |||||
| 21OmniDocBenchShanghai AI Lab (OpenDataLab) | 6 | 13 | 4 | 1 | 1 | 1 | 1 | 13 | 18/50 | 40 | |||||
| 22IFEvalGoogle | 5 | 12 | 2 | 6 | 13 | 16/50 | 38 | ||||||||
| 23ToolathlonHKUST NLP | 1 | 6 | 7 | 7 | 5 | 2 | 5 | 14 | 15/50 | 47 | |||||
| 24AA-LCRArtificial Analysis | 11 | 7 | 1 | 1 | 1 | 20 | 15/50 | 41 | |||||||
| 25OCRBenchYuliang Liu, Xiang Bai et al. | 12 | 2 | 1 | 1 | 1 | 11 | 14/50 | 28 | |||||||
| 26τ³-bench BankingSierra | 1 | 1 | 4 | 1 | 2 | 1 | 12 | 14/50 | 22 | ||||||
| 27SimpleQAGoogle DeepMind | 2 | 4 | 1 | 1 | 3 | 1 | 9 | 14/50 | 21 | ||||||
| 28OSWorldXLANG Lab (HKU) + collaborators | 6 | 11 | 13 | 11 | 2 | 2 | 7 | 1 | 11 | 13/50 | 64 | ||||
| 29Agents' Last ExamUC Berkeley RDI + 300+ experts | 5 | 4 | 9 | 2 | 1 | 5 | 10 | 13/50 | 36 | ||||||
| 30Berkeley Function Calling Leaderboard (BFCL)UC Berkeley (Gorilla) | 12 | 3 | 7 | 1 | 10 | 13/50 | 33 | ||||||||
| 31IMO-AnswerBenchGoogle DeepMind | 8 | 2 | 3 | 3 | 15 | 13/50 | 31 | ||||||||
| 32LongBenchTsinghua KEG / Zhipu | 11 | 3 | 1 | 2 | 1 | 9 | 13/50 | 27 | |||||||
| 33SWE AtlasScale AI | 1 | 2 | 1 | 10 | 13/50 | 14 | |||||||||
| 34MATH-Vision (MathVision)CUHK MMLab | 6 | 13 | 2 | 1 | 13 | 12/50 | 35 | ||||||||
| 35ProgramBenchMeta Superintelligence Labs, Stanford University and Harvard University | 2 | 8 | 2 | 2 | 2 | 1 | 10 | 12/50 | 27 | ||||||
| 36DeepSearchQAGoogle DeepMind | 5 | 4 | 2 | 13 | 12/50 | 24 | |||||||||
| 37MathVistaUCLA + Microsoft | 11 | 1 | 1 | 1 | 9 | 12/50 | 23 | ||||||||
| 38MMMUMMMU team (OSU, Waterloo and others) | 12 | 1 | 1 | 1 | 8 | 11/50 | 23 | ||||||||
| 39Video-MMEVideo-MME team (USTC, XMU, et al.) | 16 | 2 | 1 | 11 | 10/50 | 30 | |||||||||
| 40BabyVisionUniPat AI + xbench | 10 | 1 | 1 | 2 | 2 | 11 | 10/50 | 27 | |||||||
| 41Artificial Analysis Intelligence IndexArtificial Analysis | 4 | 6 | 2 | 3 | 1 | 8 | 10/50 | 24 | |||||||
| 42MultiChallengeScale AI | 9 | 4 | 1 | 1 | 8 | 10/50 | 23 | ||||||||
| 43MMStarUSTC and Shanghai AI Laboratory | 12 | 1 | 1 | 7 | 10/50 | 21 | |||||||||
| 44Vals Finance AgentVals AI | 3 | 3 | 5 | 1 | 3 | 1 | 5 | 10/50 | 21 | ||||||
| 45HallusionBenchUniversity of Maryland | 9 | 1 | 1 | 7 | 10/50 | 18 | |||||||||
| 46CritPtCritPt collaboration | 2 | 2 | 3 | 1 | 8 | 10/50 | 16 | ||||||||
| 47BLINKXingyu Fu et al. | 1 | 1 | 9 | 10/50 | 11 | ||||||||||
| 48SimpleVQAMultimodal Art Projection (M-A-P) et al. | 14 | 1 | 2 | 1 | 8 | 9/50 | 26 | ||||||||
| 49FrontierCodeCognition | 2 | 5 | 8 | 1 | 1 | 1 | 6 | 9/50 | 24 | ||||||
| 50GDP.pdfSurge AI | 2 | 3 | 4 | 1 | 3 | 5 | 9/50 | 18 | |||||||
| 51PostTrainBenchAISA group (ELLIS Institute Tübingen and MPI-IS) + University of Tübingen + Thoughtful Lab | 1 | 6 | 2 | 1 | 8 | 9/50 | 18 | ||||||||
| 52MMLUUC Berkeley (Hendrycks et al.) | 1 | 4 | 1 | 1 | 6 | 9/50 | 13 | ||||||||
| 53HealthBench ProfessionalOpenAI | 1 | 1 | 12 | 9 | 3 | 1 | 1 | 1 | 8/50 | 29 | |||||
| 54ARC-AGIARC Prize Foundation | 3 | 7 | 6 | 2 | 8 | 8/50 | 26 | ||||||||
| 55GraphWalksOpenAI | 7 | 7 | 6 | 1 | 1 | 3 | 8/50 | 25 | |||||||
| 56HMMT Nov 2025MathArena (ETH Zurich SRI Lab), problems by HMMT | 13 | 3 | 8 | 8/50 | 24 | ||||||||||
| 57ScreenSpot-ProScreenSpot-Pro authors | 11 | 1 | 3 | 2 | 2 | 1 | 3 | 8/50 | 23 | ||||||
| 58AI2DAllen Institute for AI (AI2) | 11 | 1 | 1 | 5 | 8/50 | 18 | |||||||||
| 59JobBenchUniversity of Washington + collaborators | 5 | 5 | 8 | 8/50 | 18 | ||||||||||
| 60RefCOCO / RefCOCO+ / RefCOCOgUNC Chapel Hill (Licheng Yu et al.) | 11 | 1 | 1 | 1 | 4 | 8/50 | 18 | ||||||||
| 61FrontierSWEProximal | 2 | 1 | 7 | 2 | 1 | 2 | 2 | 8/50 | 17 | ||||||
| 62OpenAI MRCROpenAI | 8 | 3 | 1 | 1 | 4 | 8/50 | 17 | ||||||||
| 63Harvey LABHarvey | 2 | 1 | 7 | 1 | 2 | 1 | 2 | 8/50 | 16 | ||||||
| 64Claw-EvalClaw-Eval team (academic) | 5 | 1 | 2 | 6 | 8/50 | 14 | |||||||||
| 65PinchBenchKilo AI | 3 | 1 | 1 | 2 | 7 | 8/50 | 14 | ||||||||
| 66Claw-Eval Pass^3Claw-Eval team | 3 | 1 | 9 | 8/50 | 13 | ||||||||||
| 67APEX-AgentsMercor | 1 | 1 | 10 | 8/50 | 12 | ||||||||||
| 68NL2Repo-BenchByteDance + M-A-P + 2077AI | 11 | 4 | 5 | 10 | 7/50 | 30 | |||||||||
| 69MMMLUOpenAI | 8 | 11 | 4 | 1 | 4 | 7/50 | 28 | ||||||||
| 70SuperGPQAByteDance Seed + M-A-P | 16 | 2 | 2 | 6 | 7/50 | 26 | |||||||||
| 71ZeroBenchUniversity of Cambridge and others | 11 | 2 | 3 | 7 | 7/50 | 23 | |||||||||
| 72MRCRGoogle DeepMind | 12 | 3 | 2 | 2 | 3 | 7/50 | 22 | ||||||||
| 73Video-MMMUNTU S-Lab (LMMs-Lab) | 1 | 13 | 6 | 7/50 | 20 | ||||||||||
| 74ChartographySurge AI | 4 | 1 | 6 | 1 | 3 | 3 | 7/50 | 18 | |||||||
| 75AA-BriefcaseArtificial Analysis | 1 | 7 | 1 | 2 | 6 | 7/50 | 17 | ||||||||
| 76OfficeQA ProDatabricks | 1 | 10 | 1 | 4 | 7/50 | 16 | |||||||||
| 77Arena-HardArena (formerly LMArena) | 1 | 2 | 1 | 9 | 7/50 | 13 | |||||||||
| 78SWE-MarathonAbundant AI | 2 | 2 | 3 | 5 | 7/50 | 12 | |||||||||
| 79ERQAGoogle DeepMind | 1 | 17 | 1 | 1 | 4 | 6/50 | 24 | ||||||||
| 80WMT24++Google | 3 | 11 | 5 | 1 | 4 | 6/50 | 24 | ||||||||
| 81MedXpertQA-MMTsinghua University (TsinghuaC3I) | 6 | 11 | 1 | 1 | 2 | 6/50 | 21 | ||||||||
| 82Terminal-Bench-ScienceStanford-led community (Terminal-Bench team + domain experts) | 4 | 5 | 5 | 4 | 6/50 | 18 | |||||||||
| 83MMVUYale NLP | 10 | 1 | 6 | 6/50 | 17 | ||||||||||
| 84WideSearchByteDance Seed | 7 | 10 | 6/50 | 17 | |||||||||||
| 85BioMysteryBenchAnthropic | 2 | 1 | 10 | 3 | 6/50 | 16 | |||||||||
| 86CountBenchGoogle Research | 11 | 1 | 4 | 6/50 | 16 | ||||||||||
| 87SkillsBenchBenchFlow + 105 domain experts | 7 | 1 | 6 | 6/50 | 14 | ||||||||||
| 88BrowseComp-ZHResearchers from Peking University, Mindverse AI, and other groups | 4 | 1 | 4 | 6/50 | 9 | ||||||||||
| 89GSM8KOpenAI | 1 | 3 | 1 | 2 | 1 | 1 | 6/50 | 9 | |||||||
| 90HumanEvalOpenAI | 1 | 3 | 2 | 3 | 6/50 | 9 | |||||||||
| 91IMOInternational Mathematical Olympiad, hosted by the Australian Maths Trust | 1 | 1 | 2 | 4 | 6/50 | 8 | |||||||||
| 92ScreenSpotOS-Atlas team (Shanghai AI Lab and others) | 1 | 1 | 1 | 2 | 2 | 6/50 | 7 | ||||||||
| 93RealWorldQAxAI | 17 | 1 | 4 | 5/50 | 22 | ||||||||||
| 94Codeforces ratingCodeforces | 7 | 4 | 1 | 3 | 1 | 5/50 | 16 | ||||||||
| 95HealthBench HardOpenAI | 1 | 11 | 1 | 2 | 5/50 | 15 | |||||||||
| 96VITA-BenchMeituan LongCat team | 11 | 4 | 5/50 | 15 | |||||||||||
| 97OfficeQADatabricks | 1 | 10 | 1 | 2 | 5/50 | 14 | |||||||||
| 98V*Penghao Wu and Saining Xie | 8 | 1 | 1 | 3 | 5/50 | 13 | |||||||||
| 99DRACOPerplexity | 8 | 4 | 5/50 | 12 | |||||||||||
| 100FLEURSGoogle | 5 | 4 | 1 | 2 | 5/50 | 12 | |||||||||
| 101MathArena (Apex, IMO, USAMO)ETH Zurich SRI Lab | 2 | 4 | 6 | 5/50 | 12 | ||||||||||
| 102Terminal-Bench HardArtificial Analysis | 4 | 1 | 4 | 5/50 | 9 | ||||||||||
| 103WorldVQAMoonshot AI | 2 | 1 | 6 | 5/50 | 9 | ||||||||||
| 104FrontierScienceOpenAI | 1 | 7 | 5/50 | 8 | |||||||||||
| 105InfographicVQAComputer Vision Center (UAB) and IIIT Hyderabad | 4 | 1 | 1 | 2 | 5/50 | 8 | |||||||||
| 106MATHUC Berkeley (Hendrycks et al.) | 3 | 1 | 2 | 2 | 5/50 | 8 | |||||||||
| 107WebArena-VerifiedServiceNow | 2 | 1 | 1 | 1 | 2 | 5/50 | 7 | ||||||||
| 108AdvancedIFMeta Superintelligence Labs | 2 | 4 | 5/50 | 6 | |||||||||||
| 109MMSI-BenchShanghai AI Lab (OpenRobotLab) and collaborators | 1 | 5 | 5/50 | 6 | |||||||||||
| 110WebDev ArenaArena (formerly LMArena) | 2 | 1 | 1 | 2 | 5/50 | 6 | |||||||||
| 111ChartQAYork University and Nanyang Technological University | 1 | 1 | 1 | 2 | 5/50 | 5 | |||||||||
| 112DocVQACVC, Universitat Autònoma de Barcelona and IIIT Hyderabad | 1 | 1 | 1 | 2 | 5/50 | 5 | |||||||||
| 113LogicVistaUCLA | 1 | 1 | 3 | 5/50 | 5 | ||||||||||
| 114MathVerseCUHK MMLab and Shanghai AI Laboratory | 1 | 4 | 5/50 | 5 | |||||||||||
| 115PaperBenchOpenAI | 1 | 4 | 5/50 | 5 | |||||||||||
| 116LVBenchZhipu AI / Tsinghua | 4 | 14 | 4 | 4/50 | 22 | ||||||||||
| 117C-EvalSJTU + Tsinghua + Edinburgh | 14 | 1 | 2 | 1 | 4/50 | 18 | |||||||||
| 118MMLU-ProXUniversity of Tokyo (Li Lab) and collaborators | 11 | 5 | 1 | 1 | 4/50 | 18 | |||||||||
| 119DynaMathUniversity of Illinois Urbana-Champaign and UC Berkeley | 11 | 4 | 4/50 | 15 | |||||||||||
| 120EmbSpatialBenchFudan University | 11 | 3 | 4/50 | 14 | |||||||||||
| 121MMBenchShanghai AI Laboratory (OpenCompass) | 11 | 1 | 2 | 4/50 | 14 | ||||||||||
| 122CursorBenchCursor | 1 | 7 | 2 | 3 | 4/50 | 13 | |||||||||
| 123RULERNVIDIA | 4 | 6 | 1 | 1 | 4/50 | 12 | |||||||||
| 124OJBenchBeijing University of Posts and Telecommunications, Tsinghua University, and others | 6 | 4 | 4/50 | 10 | |||||||||||
| 125BBEH micro averageGoogle DeepMind | 6 | 1 | 2 | 4/50 | 9 | ||||||||||
| 126GDPvalOpenAI | 5 | 1 | 1 | 2 | 4/50 | 9 | |||||||||
| 127MATH-500OpenAI | 3 | 3 | 3 | 4/50 | 9 | ||||||||||
| 128VoiceBenchCUHK-Shenzhen | 3 | 2 | 4 | 4/50 | 9 | ||||||||||
| 129Artificial Analysis Coding Agent IndexArtificial Analysis | 1 | 4 | 1 | 1 | 4/50 | 7 | |||||||||
| 130Beyond AIMEByteDance Seed | 1 | 6 | 4/50 | 7 | |||||||||||
| 131HellaSwagUniversity of Washington + AllenAI | 3 | 1 | 2 | 1 | 4/50 | 7 | |||||||||
| 132MGSMGoogle Research | 1 | 3 | 2 | 1 | 4/50 | 7 | |||||||||
| 133MMAUUniversity of Maryland | 3 | 1 | 3 | 4/50 | 7 | ||||||||||
| 134USAMOMathematical Association of America (contest), MathArena (evaluation) | 4 | 1 | 2 | 4/50 | 7 | ||||||||||
| 135We-MathBeijing University of Posts and Telecommunications | 4 | 3 | 4/50 | 7 | |||||||||||
| 136Audio MCScale AI | 1 | 1 | 1 | 3 | 4/50 | 6 | |||||||||
| 137BigCodeBenchBigCode project | 1 | 2 | 2 | 1 | 4/50 | 6 | |||||||||
| 138CL-benchTencent Hunyuan (with Fudan University authors) | 1 | 5 | 4/50 | 6 | |||||||||||
| 139LMArena (Elo)Arena (formerly LMArena) | 3 | 1 | 1 | 1 | 4/50 | 6 | |||||||||
| 140LongVideoBenchNTU S-Lab et al. | 6 | 4/50 | 6 | ||||||||||||
| 141MBPPGoogle Research | 3 | 3 | 4/50 | 6 | |||||||||||
| 142PerceptionBench (Moonshot)Moonshot AI | 1 | 1 | 4 | 4/50 | 6 | ||||||||||
| 143ResearchRubricsScale AI | 6 | 4/50 | 6 | ||||||||||||
| 144Big Bench AudioArtificial Analysis | 1 | 2 | 2 | 4/50 | 5 | ||||||||||
| 145BIG-Bench HardSuzgun et al. (Stanford University and Google Research) | 1 | 2 | 2 | 4/50 | 5 | ||||||||||
| 146ChartQA Provis-nlp (Masry et al.) | 1 | 4 | 4/50 | 5 | |||||||||||
| 147Chinese SimpleQAAlibaba (Taobao & Tmall) | 1 | 2 | 2 | 4/50 | 5 | ||||||||||
| 148CVBenchNew York University (Cambrian-1 team) | 2 | 1 | 2 | 4/50 | 5 | ||||||||||
| 149FinSearchCompByteDance Seed | 5 | 4/50 | 5 | ||||||||||||
| 150LibriSpeechPanayotov et al. (Johns Hopkins University) | 2 | 3 | 4/50 | 5 | |||||||||||
| 151LiveCodeBench ProLiveCodeBench Pro team | 1 | 1 | 2 | 1 | 4/50 | 5 | |||||||||
| 152WildClawBenchInternLM (Shanghai AI Lab) | 1 | 1 | 3 | 4/50 | 5 | ||||||||||
| 153Online-Mind2WebOhio State University (OSU NLP Group) | 1 | 1 | 1 | 1 | 4/50 | 4 | |||||||||
| 154PutnamBenchUT Austin (Tsoukalas et al.) | 1 | 1 | 1 | 1 | 4/50 | 4 | |||||||||
| 155WebVoyagerWebVoyager team (He et al., ACL 2024) | 1 | 1 | 1 | 1 | 4/50 | 4 | |||||||||
| 156MMLU-ReduxUniversity of Edinburgh | 16 | 1 | 1 | 3/50 | 18 | ||||||||||
| 157INCLUDEEPFL and Cohere For AI | 11 | 4 | 1 | 3/50 | 16 | ||||||||||
| 158MLVUBAAI + partners | 14 | 2 | 3/50 | 16 | |||||||||||
| 159BenchCADBenchCAD team | 6 | 7 | 1 | 3/50 | 14 | ||||||||||
| 160MMLongBench-DocYubo Ma et al. | 10 | 2 | 1 | 3/50 | 13 | ||||||||||
| 161SWE-bench MultimodalSWE-bench team (Princeton/Stanford) | 1 | 10 | 2 | 3/50 | 13 | ||||||||||
| 162MLE-benchOpenAI | 2 | 6 | 2 | 3/50 | 10 | ||||||||||
| 163VLMs Are BlindAuburn University and University of Alberta | 8 | 2 | 3/50 | 10 | |||||||||||
| 164MCPMarkeval-sys (MCPMark team) | 5 | 4 | 3/50 | 9 | |||||||||||
| 165Multi-IFMeta | 7 | 2 | 3/50 | 9 | |||||||||||
| 166Vending-BenchAndon Labs | 2 | 5 | 2 | 3/50 | 9 | ||||||||||
| 167Claw-Eval-MMClaw-Eval team | 6 | 2 | 3/50 | 8 | |||||||||||
| 168LAB-BenchEdison Scientific (FutureHouse) | 3 | 1 | 3 | 3/50 | 7 | ||||||||||
| 169Vision2WebZ.ai (zai-org) | 4 | 2 | 1 | 3/50 | 7 | ||||||||||
| 170AGIEvalMicrosoft | 3 | 2 | 1 | 3/50 | 6 | ||||||||||
| 171Seal-0Virginia Tech | 4 | 2 | 3/50 | 6 | |||||||||||
| 172Vals IndexVals AI | 3 | 3 | 3/50 | 6 | |||||||||||
| 173Daily-OmniZiwei Zhou, Rui Wang et al. | 3 | 1 | 1 | 3/50 | 5 | ||||||||||
| 174FrontierScience-ResearchOpenAI | 2 | 1 | 2 | 3/50 | 5 | ||||||||||
| 175Harvey LAB-AA v1.1Artificial Analysis | 2 | 1 | 2 | 3/50 | 5 | ||||||||||
| 176MLS-BenchMLS-Bench authors (Bohan Lyu et al.) | 2 | 3 | 3/50 | 5 | |||||||||||
| 177MotionBenchZhipu AI and Tsinghua University (THUDM) | 5 | 3/50 | 5 | ||||||||||||
| 178WinoGrandeAllenAI | 3 | 1 | 1 | 3/50 | 5 | ||||||||||
| 179Bird-SQLHKU and Alibaba DAMO Academy | 1 | 2 | 1 | 3/50 | 4 | ||||||||||
| 180EnterpriseOps-GymServiceNow | 1 | 3 | 3/50 | 4 | |||||||||||
| 181Full-Duplex-BenchGuan-Ting Lin, Hung-yi Lee and collaborators (National Taiwan University and others) | 1 | 2 | 1 | 3/50 | 4 | ||||||||||
| 182Global MMLU LiteCohere Labs | 1 | 3 | 3/50 | 4 | |||||||||||
| 183IMO-ProofBenchGoogle DeepMind (Superhuman Reasoning team) | 2 | 2 | 3/50 | 4 | |||||||||||
| 184Seed-TTS-evalByteDance (Seed-TTS team) | 2 | 2 | 3/50 | 4 | |||||||||||
| 185SpreadsheetBenchRenmin University (RUCKBReasoning) | 2 | 2 | 3/50 | 4 | |||||||||||
| 186SWE-Bench Pro V2 HardScale AI (co-developed with Reflection) | 4 | 3/50 | 4 | ||||||||||||
| 187VisuLogicShanghai AI Laboratory with several universities | 4 | 3/50 | 4 | ||||||||||||
| 188WebArenaCarnegie Mellon University | 2 | 1 | 1 | 3/50 | 4 | ||||||||||
| 189WenetSpeechNorthwestern Polytechnical University (ASLP@NPU), Mobvoi and others | 2 | 2 | 3/50 | 4 | |||||||||||
| 190$OneMillion-BenchHumanlaya + BIGAI + xbench | 1 | 2 | 3/50 | 3 | |||||||||||
| 191AMO-BenchMeituan LongCat Team | 1 | 2 | 3/50 | 3 | |||||||||||
| 192ArtifactsBenchTencent Hunyuan | 1 | 2 | 3/50 | 3 | |||||||||||
| 193BrowseComp-VLAlibaba Tongyi Lab | 1 | 1 | 1 | 3/50 | 3 | ||||||||||
| 194Legal Research BenchVals AI | 1 | 2 | 3/50 | 3 | |||||||||||
| 195MindCubeMLL Lab, Northwestern University, with Stanford, NYU, and University of Washington | 1 | 2 | 3/50 | 3 | |||||||||||
| 196MM-BrowseCompByteDance Seed and collaborators (Li et al.) | 1 | 2 | 3/50 | 3 | |||||||||||
| 197MM-IFEvalMM-IFEngine team (Ding et al.) | 1 | 2 | 3/50 | 3 | |||||||||||
| 198MuirBenchFei Wang, Xingyu Fu et al. | 1 | 2 | 3/50 | 3 | |||||||||||
| 199olmOCR-BenchAllen Institute for AI | 3 | 3/50 | 3 | ||||||||||||
| 200PhyXHui Shen et al. | 1 | 2 | 3/50 | 3 | |||||||||||
| 201PresentBenchPresentBench team | 1 | 2 | 3/50 | 3 | |||||||||||
| 202SpreadsheetBench 2Renmin University of China (RUCKBReasoning) | 3 | 3/50 | 3 | ||||||||||||
| 203TextVQAFacebook AI Research | 1 | 1 | 1 | 3/50 | 3 | ||||||||||
| 204TriviaQAUniversity of Washington | 1 | 1 | 1 | 3/50 | 3 | ||||||||||
| 205ViewSpatial-BenchDingming Li, Yueting Zhuang et al. | 1 | 2 | 3/50 | 3 | |||||||||||
| 206WorkSpaceBenchSJTU + ByteDance (OpenDataBox) | 1 | 2 | 3/50 | 3 | |||||||||||
| 207WritingBenchAlibaba Group (X-PLUG) | 3 | 3/50 | 3 |
How this is counted
A release is one launch post. When a lab put out several models under one post, that's one release. The system cards, model cards and technical reports that came with a post count toward the same release. A benchmark counts once per release, however many settings, subsets or documents report it, and with or without a score.
Versions of one test count as one benchmark in the graphics, so Terminal-Bench 2.0 and 3.0 are both Terminal-Bench. Tests with their own task sets stay apart, like SWE-bench Verified, SWE-Bench Pro and SWE-bench Multilingual. The lookup can show every version on its own.
A benchmark's color and shape come from its area, the same eight areas the rest of the almanac uses. Safety, alignment, bio and cyber evaluations are left out until you switch them on. Most come from system cards, and only a few labs publish those.
The quarterly chart counts 2024 as one column, because we have only 31 launch posts with scores from that year. The spread chart only shows benchmarks released since 2025, so the first lab we saw name one is the first lab that did.
The charts over time count by half year from 2024. We have read far fewer launch posts from 2024 than from later years (16 from the first half of 2024, 142 from the busiest half year since), so the earliest points rest on fewer releases. Hover a point to see how many. The latest half year isn't over yet, so its point is hollow.
A benchmark stays in launch posts until no release has named any version of it for 6 months. We can only tell that for benchmarks last named before May 21st, 2026, so a benchmark named since then counts as followed up to that day, and one first named since then doesn't count yet. Each line is a survival estimate, as for saturation on the evals page, and stops when fewer than 15 benchmarks are left to follow. The first year we read is 2024, so benchmarks labs had named before then and named again in 2024 count as first named in 2024.
"Frontier models only" counts just the releases of a model that ranked in the top ten on Epoch AI's Capabilities Index among the models out the day it was released, from any lab. It uses Epoch AI's index, so a model Epoch hasn't scored doesn't count. Both switches can be on at once.
"Frontier labs only" counts just the releases of the nine labs whose best 2026 models score highest on Epoch AI's Capabilities Index, as of November 2026. Every count on the page then compares those labs with each other, and the spread chart asks how fast a benchmark reached them.
We only count what labs reported themselves. Leaderboard and Epoch AI scores don't count here. Every number links back to its source on the benchmark's page and the model's page, and the Model cards page lists every post we read.
More insights
-
How are evals changing?
Agent tasks made up 6% of the new benchmarks in 2023 and 48% in 2026. Multiple-choice tests fell from 24% to 5%.
-
How fast is the frontier moving?
The best open model today, Kimi K3.1 Pro, scores what the best closed model scored seven months ago. Since 2024 the median gap has been six months.
-
How do new models get tested?
Of the benchmarks labs named in launch posts since 2024, an official leaderboard or Epoch AI has scored the same model on 14%.