Skip to Sidebar Skip to Content
Company tracker 3 of 10 · 10 companies, one lens

Google

Google and Google DeepMind, makers of the Gemini models, tracked through the same six lenses as every company on AI Adoption.
Its best model, Gemini 4 Argon, scores 166.4 on the Epoch Capabilities Index, #3 of 10, 2.4 behind Anthropic. In March 2026, 20.9% of US adults said they had used Google Gemini in the past week. It released 15 notable models in 2025, #2 of 10 by count. It owns AI chips equal to about 5.47 million H100s.
Follow GoogleCompare with the others
48 ledger entries23 charts name GoogleLatest entry Sep 2026Checked 20 Oct 2026

Google through six lenses

How the lenses work →
The same six lenses, in the same order, on every company page. Each shows one comparable number where a source measures several companies the same way, its rank, and when it last changed.
01Capability#3 of 10
How good are its best models?
166.4
Epoch Capabilities Index score of its best model, Gemini 4 Argon (Sep 2026)
Anthropic168.8OpenAI168.8Google166.4Meta161.8
Epoch combines many benchmarks into one scale. 90% interval for this model: 163.2 to 171.5.
Text Arena, Mar 2026: 1,494, #3 of 8. Arena chart →
Independent measureChanged Sep 2026Epoch AI data →
02Output#2 of 10
How much does it ship?
15
Notable AI models released in 2025
OpenAI21Google15Alibaba11Anthropic7
194 notable models from 2014 to 2025 and 6 in 2026 so far. Model record →
Independent measureChanged 2025View chart →
03Money
How much does it earn?
Not measured yet
No AI revenue figure for Google in Epoch AI's company data
Epoch AI tracks revenue for AI-first companies such as OpenAI, Anthropic, Mistral AI, DeepSeek. about 6,000 staff at Google DeepMind (Aug 2025).
How a lens gets a number →
04Compute#4 of 10
How much compute and energy goes into its AI?
5×10²⁵
FLOP: its largest training run with a published estimate, Gemini 1.0 Ultra (Dec 2023)
OpenAI9.4×10²⁶xAI5×10²⁶Meta5.2×10²⁵Google5×10²⁵
Largest estimated training cost: $31M (Gemini 1.0 Ultra), Epoch AI. Model record →
Ranked among the 10 tracked companies. Only runs with a published estimate count, so a company that discloses less can rank lower.
Owns AI chips equal to about 5.47M H100s (Mar 2026), #1 of 7 named owners; had about 1.58M H100-equivalents available at the end of 2025, #2 of 5 labs. Epoch AI data →
Independent estimateChanged Dec 2023View chart →
05Use#2 of 7
Who uses it, and for what?
20.9%
of US adults said they used Google Gemini in the past week, March 2026
OpenAI31.0%Google20.9%Microsoft10.5%Meta7.5%
Epoch AI with Ipsos, a probability-based sample of 2,021 US adults; 49.2% used no AI service. Among employed adults in July 2026: 24.2%.
92% of symptom and common health searches analysed showed an AI Overview. Chart →
Independent measureChanged Mar 2026Epoch AI data →
06Safety & policy#3 of 9
How reliable is it, and how is it governed?
50%
hallucination rate of its most reliable model on AA-Omniscience, Gemini 3.1 Pro Preview
xAI22%Anthropic26%Google50%DeepSeek82%
Lower is better. Ranked among the tracked companies; the test covers 26 models. One reliability test: safety frameworks, evaluations and regulatory actions are logged as entries.
Independent measureChanged 2026View chart →

The ledger

What gets an entry →
An entry goes in only if it changes a number or a position on one of the six lenses. If it moves nothing, it is not news here. Newest first.

  1. September 2026

    1. 30 Sep 2026CapabilityNewGemini 4 Argon scored 166.4 on the Epoch Capabilities Index, its best model#3 of the 10 tracked companies' best models; Anthropic's best scored 168.8. Epoch's 90% interval: 163.2 to 171.5.Independent measureEpoch AI, Benchmarking HubEpoch AI data →#3

  2. July 2026

    1. Jul 2026UseNew24.2% of employed US adults said they had used Google Gemini in the past weekProbability-based sample of 1,103 employed US adults, asked about work and personal use.Independent measureEpoch AI and IpsosEpoch AI data →

  3. March 2026

    1. Mar 2026ComputeNewOwned AI chips equal to about 5.47 million Nvidia H100sEpoch's 90% range: 4.39M to 6.63M. #1 of the 7 owners Epoch lists by name.Independent estimateEpoch AI, AI Chip OwnersEpoch AI data →5.47M
    2. Mar 2026UseNew20.9% of US adults said they had used Google Gemini in the past week#2 of the 7 services asked about that belong to tracked companies. 49.2% had used no AI service. Probability-based sample of 2,021 US adults.Independent measureEpoch AI and IpsosEpoch AI data →#2

  4. 2026 tests

    1. 2026 testSafety & policyNewIts most reliable model on AA-Omniscience, Gemini 3.1 Pro Preview, had a 50% hallucination rate#3 of 9 tracked companies; xAI's best was 22%. Rates across the 26 models tested ran from 22% to 94%.Independent measureArtificial AnalysisView chart →#3 of 9

  5. March 2026

    1. Mar 2026CapabilityNewFell to #3 of 8 on the Text Arena, 9 points behind AnthropicAnthropic led at 1,503. The top four labs sat within 22 Arena points.Independent measureArenaView chart →#2 → #3

  6. February 2026

    1. Feb 2026CapabilityNewGemini 3 Deep Think scored 84.6% on ARC-AGI-2, 7.5 points ahead of the next systemThe next system was Gemini 3.1 Pro at 77.1%. Its earlier preview version scored 45.1%.Independent measureARC-AGI-2 LeaderboardView chart →#1
    2. Feb 2026CapabilityNewGemini 3.1 Pro Preview scored 88.2% on MMMU, 0.4 points short of the best human expertCollege-level questions combining text and images. The 14th model, o3, scored 80.4%.Independent measureVals.aiView chart →#1
    3. Feb 2026CapabilityNewgemini-3-pro led the Vision Arena at 1,288; Gemini models held four of the top six places10 points ahead of gemini-3.1-pro-preview.Independent measureArenaView chart →#1
    4. Feb 2026CapabilityNewFell to #2 of 8 on the Text Arena, 4 points behind AnthropicAnthropic led at 1,504. First to fourth spanned 23 points.Independent measureArenaView chart →#1 → #2

  7. Early 2026 leaderboards

    1. Early 2026CapabilityNewGemini-3.1-Pro led MMLU-Pro at 91.2%The top 15 models were within 4.2 points; Grok-4 was 15th at 87.0%.Independent measureMMLU-Pro LeaderboardView chart →#1
    2. Early 2026CapabilityNewGemini 3.1 Pro Preview led LegalBench at 87.4%Just ahead of Gemini 3 Pro at 87%. The top 15 span only 4.0 points.Independent measureVals.aiView chart →#1
    3. Early 2026CapabilityNewGemini 3.1 Pro Preview led MortgageTax at 69.4%; no model reached 70%The top 15 models sat within 3.5 points.Independent measureVals.aiView chart →#1

  8. End of 2025

    1. End of 2025ComputeNewHad compute equal to about 1.58 million H100s available at the end of 2025Owned or rented. Epoch's 90% range: 1.01M to 2.55M. #2 of 5 labs.Independent estimateEpoch AI, AI Chip UsersEpoch AI data →1.58M

  9. 2014–2025

    1. 2014–2025OutputNewGoogle has released 194 notable AI models since 2014, the most of any organisationMore than twice Meta's 87. OpenAI is third with 61.Independent measureEpoch AIView chart →#1

  10. 2025, full year

    1. 2025OutputNewReleased 15 notable AI models in 2025, #2 by countOpenAI 21, Alibaba 11. Epoch AI counts a model as notable if it was state of the art, highly cited or historically significant.Independent measureEpoch AIView chart →#2
    2. 2025CapabilityNewgemini-3-pro-preview set the best GPQA Diamond score, 93%, 11.8 points above expert validatorsThe best score in 2023 was 36%, from GPT-4.Independent measureEpoch AIView chart →#1
    3. 2025CapabilityNewGemini 3 Pro set the best score on Humanity's Last Exam, 38.3%The best in 2024 was 8%, from o1.Independent measureCenter for AI Safety et al.View chart →#1

  11. Q3 2025

    1. Q3 2025UseNewGoogle's share of downloads of the top Hugging Face models fell to 0.3%It was 42% in 2020.Independent measureLongpre et al., Hugging FaceView chart →−41.7 pts

  12. August 2025

    1. Aug 2025CapabilityNewTook first place on the Text Arena with 1,461 pointsAnthropic was second at 1,458. First to fourth spanned 30 points.Independent measureArenaView chart →#2 → #1

  13. July 2025

    1. Jul 2025CapabilityNewFell to #2 of 8 on the Text Arena, 20 points behind OpenAIOpenAI led at 1,482. First to fourth spanned 50 points.Independent measureArenaView chart →#1 → #2

  14. 2025 studies

    1. 2025 studyCapabilityNewGemini 2.5 Pro scored 38.8% on PaperArena as a multi-agent system, the best of nine modelsThe AI Index puts PhD experts at 83.5%.Independent measureWang et al.View chart →#1 of 9
    2. 2025 studyCapabilityNewGemini-2.0 read analogue clocks exactly right 22.6% of the time, the best of seven modelsGPT-4o managed 8.1% and Qwen2-VL-7B none.Independent measureSaxena et al.View chart →#1 of 7
    3. 2025 studyUseNewAI Overviews appeared on 84% to 92% of the health-related Google searches analysedSymptom searches and common health questions returned one 92% of the time.Independent measureRAISE HealthView chart →

  15. May 2025

    1. May 2025CapabilityNewTook first place on the Text Arena with 1,492 pointsOpenAI was second at 1,458. First to fourth spanned 80 points.Independent measureArenaView chart →#6 → #1

  16. April 2025

    1. Apr 2025CapabilityNewFell to #6 of 8 on the Text Arena, 92 points behind OpenAIOpenAI led at 1,461. First to fourth spanned 70 points.Independent measureArenaView chart →#5 → #6

  17. March 2025

    1. Mar 2025CapabilityNewFell to #5 of 8 on the Text Arena, 71 points behind OpenAIOpenAI led at 1,441. First to fourth spanned 46 points.Independent measureArenaView chart →#4 → #5

  18. December 2024

    1. Dec 2024CapabilityNewFell to #4 of 8 on the Text Arena, 56 points behind OpenAIOpenAI led at 1,407. First to fourth spanned 56 points.Independent measureArenaView chart →#3 → #4

  19. October 2024

    1. Oct 2024CapabilityNewFell to #3 of 8 on the Text Arena, 31 points behind OpenAIOpenAI led at 1,384. First to fourth spanned 50 points.Independent measureArenaView chart →#2 → #3

  20. September 2024

    1. Sep 2024CapabilityNewRose to #2 of 8 on the Text Arena, 35 points behind OpenAIOpenAI led at 1,387. First to fourth spanned 53 points.Independent measureArenaView chart →#5 → #2

  21. August 2024

    1. Aug 2024CapabilityNewFell to #5 of 8 on the Text Arena, 13 points behind OpenAIOpenAI led at 1,341. First to fourth spanned 10 points.Independent measureArenaView chart →#3 → #5

  22. May 2024

    1. 8 May 2024OutputNewAlphaFold 3 released, trained with an estimated 4.1×10²² FLOPAbout 51× its previous model in this data, GenCast (May 2024).Independent estimateEpoch AIView chart →4.1×10²²
    2. May 2024CapabilityNewRose to #3 of 7 on the Text Arena, 23 points behind OpenAIOpenAI led at 1,340. First to fourth spanned 70 points.Independent measureArenaView chart →#6 → #3
    3. 1 May 2024OutputNewGenCast released, trained with an estimated 8.2×10²⁰ FLOPLess training compute than FunSearch (3.9×10²³).Independent estimateEpoch AIView chart →8.2×10²⁰

  23. April 2024

    1. Apr 2024CapabilityNewFell to #6 of 7 on the Text Arena, 109 points behind OpenAIOpenAI led at 1,338. First to fourth spanned 91 points.Independent measureArenaView chart →#5 → #6

  24. March 2024

    1. Mar 2024CapabilityNewFell to #5 of 7 on the Text Arena, 113 points behind AnthropicAnthropic led at 1,333. First to fourth spanned 95 points.Independent measureArenaView chart →#4 → #5

  25. February 2024

    1. 15 Feb 2024OutputNewGemini 1.5 Pro released, marked a frontier model in Epoch AI's databaseIndependent estimateEpoch AIModel record →
    2. Feb 2024CapabilityNewFell to #4 of 6 on the Text Arena, 113 points behind OpenAIOpenAI led at 1,329. First to fourth spanned 113 points.Independent measureArenaView chart →#3 → #4

  26. January 2024

    1. Jan 2024CapabilityNewFell to #3 of 6 on the Text Arena, 117 points behind OpenAIOpenAI led at 1,334. First to fourth spanned 145 points.Independent measureArenaView chart →#2 → #3

  27. December 2023

    1. 14 Dec 2023OutputNewFunSearch released, trained with an estimated 3.9×10²³ FLOPLess training compute than Gemini 1.0 Ultra (5×10²⁵).Independent estimateEpoch AIView chart →3.9×10²³
    2. 6 Dec 2023OutputNewGemini 1.0 Ultra released, trained with an estimated 5×10²⁵ FLOPAbout 2,381× its previous model in this data, GraphCast (Nov 2023).Independent estimateEpoch AIView chart →5×10²⁵
    3. 6 Dec 2023ComputeNewTraining Gemini 1.0 Ultra drew an estimated 19 MW of powerThe most in this data is Grok 3, at 110 MW.Independent estimateEpoch AIView chart →19 MW
    4. Dec 2023CapabilityNewRose to #2 of 5 on the Text Arena, 95 points behind OpenAIOpenAI led at 1,312. First to fourth spanned 163 points.Independent measureArenaView chart →#4 → #2

  28. November 2023

    1. 14 Nov 2023OutputNewGraphCast released, trained with an estimated 2.1×10²² FLOPLess training compute than PaLM 2 (7.3×10²⁴).Independent estimateEpoch AIView chart →2.1×10²²

  29. October 2023

    1. Oct 2023CapabilityNewFell to #4 of 5 on the Text Arena, 184 points behind OpenAIOpenAI led at 1,314. First to fourth spanned 184 points.Independent measureArenaView chart →#3 → #4

  30. August 2023

    1. Aug 2023CapabilityNewFell to #3 of 3 on the Text Arena, 197 points behind OpenAIOpenAI led at 1,313.Independent measureArenaView chart →#2 → #3

  31. May 2023

    1. 10 May 2023OutputNewPaLM 2 released, trained with an estimated 7.3×10²⁴ FLOPIndependent estimateEpoch AIView chart →7.3×10²⁴
    2. May 2023CapabilityNewPlaced #2 of 3 on the Text Arena with 1,117 points, its first month in this seriesOpenAI led at 1,322.Independent measureArenaView chart →#2

Google's model record

Download the CSV →
Every notable model Epoch AI has logged for Google: 217 in all, 194 from 2014 to 2025 and 6 in 2026 so far. Training compute, cost and hardware are Epoch's estimates, shown only where it has one.
Notable models, 2014–2025194#1 of 10 companies · Epoch AI's count, as in chart 1.1.7
Frontier models33#1 of 10 companies · marked frontier in Epoch AI's database
Open weights14%8 of 57 notable models since 2023
Largest estimated training cost$31M#4 of 10 companies · Gemini 1.0 Ultra (Dec 2023), 2023 dollars
Largest training cluster57,000#3 of 8 companies · Google TPU v4 · Gemini 1.0 Ultra
Notable models per year* 2026 to 19 Oct 2026
2014: 6620142015: 662016: 141420162017: 16162018: 141420182019: 19192020: 212120202021: 22222022: 252520222023: 25252024: 111120242025: 15152026: 662026*
Notable models since 2023, newest first57 models
ReleasedModelAccessTraining computeEst. costTraining hardware
30 Sep 2026Gemini 4 ArgonAPI or hosted access———
3 Sep 2026WeatherNext 3API or hosted access2.1×10²³——
2 Sep 2026Gemini 3.8 FlashAPI or hosted access———
19 May 2026Gemini 3.5 FlashAPI or hosted access———
15 Apr 2026Gemini Flash 3.1 TTSAPI or hosted access———
3 Mar 2026Gemini 3.0 Flash-liteAPI or hosted access———
4 Dec 2025SIMA 2Unreleased———
20 Nov 2025Gemini 3 Pro Image (Nano Banana Pro)API or hosted access———
18 Nov 2025Gemini 3 ProAPI or hosted access——Google TPU v7 Ironwood
18 Nov 2025Gemini 3 Deep ThinkAPI or hosted access———
Show all 57 models since 2023
15 Oct 2025Veo 3.1API or hosted access———
25 Sep 2025Gemini Robotics-ER 1.5API or hosted access——Google TPU v4
26 Aug 2025Gemini 2.5 Flash Image (Nano Banana)API or hosted access———
1 Aug 2025Gemini 2.5 Deep ThinkAPI or hosted access———
14 Jul 2025Gemini EmbeddingAPI or hosted access———
12 Jun 2025FGNUnreleased9.6×10²¹—Google TPU v5p
5 Jun 2025Gemini 2.5 Pro (Jun 2025)API or hosted access———
21 May 2025Veo 3API or hosted access———
20 May 2025Imagen 4API or hosted access———
6 May 2025Gemini 2.5 Pro (May 2025)API or hosted access———
25 Mar 2025Gemini 2.5 Pro (Mar 2025)API or hosted access———
16 Dec 2024Veo 2API or hosted access———
11 Dec 2024Gemini 2.0 ProAPI or hosted access———
15 Nov 2024Gemini-Exp-1114API or hosted access———
5 Sep 2024AlphaProteoUnreleased———
7 Aug 2024Table Tennis AgentUnreleased———
13 Jun 2024OpenVLAOpen weights1.1×10²³—64 NVIDIA A100
20 May 2024Octo-BaseOpen weights5.8×10²⁰—128 Google TPU v4
8 May 2024AlphaFold 3Open weights4.1×10²²—256 NVIDIA A100
1 May 2024GenCastOpen weights8.2×10²⁰—32 Google TPU v5e
15 Feb 2024Gemini 1.5 Pro frontierAPI or hosted access——Google TPU v4
17 Jan 2024AlphaGeometryOpen weights——Google TPU v3
23 Dec 2023GQA-8-XXLUnreleased3.5×10²²—Google TPU v3
19 Dec 2023Gemini Nano-2Unreleased——Google TPU v5e
19 Dec 2023Gemini Nano-1API or hosted access——Google TPU v5e
14 Dec 2023FunSearchOpen weights3.9×10²³——
11 Dec 2023W.A.L.TUnreleased———
6 Dec 2023Gemini 1.0 Ultra frontierAPI or hosted access5×10²⁵$31M57,000 Google TPU v4
6 Dec 2023Gemini 1.0 ProAPI or hosted access——Google TPU v4
29 Nov 2023GNoME for crystal discoveryUnreleased——4 Google TPU v4
14 Nov 2023GraphCastOpen weights2.1×10²²——
3 Nov 2023RT-TrajectoryUnreleased———
17 Oct 2023PaLI-3Unreleased———
13 Oct 2023RT-2-XUnreleased———
22 Sep 2023AlphaMissenseUnreleased———
28 Jul 2023RT-2Unreleased———
26 Jul 2023AudioLMUnreleased3.9×10¹⁸—Google TPU v4
29 May 2023PaLI-XUnreleased5.6×10²³——
16 May 2023Med-PaLM 2Unreleased———
10 May 2023PaLM 2 frontierAPI or hosted access7.3×10²⁴$5MGoogle TPU v4
26 Apr 2023Agile Soccer RobotUnreleased———
27 Mar 2023SigLIP 400MOpen weights4.9×10²¹—32 Google TPU v4
6 Mar 2023PaLM-EUnreleased———
13 Feb 2023BASIC-L + LionUnreleased———
10 Feb 2023ViT-22BUnreleased1.9×10²³$286K1,024 Google TPU v4
26 Jan 2023MusicLMUnreleased———
10 Jan 2023DreamerV3Unreleased2.2×10²⁰—16 NVIDIA V100
Data: Epoch AI, Data on AI Models, snapshot 19 Oct 2026, CC BY 4.0. Organisations are grouped into companies as in chart 1.1.7 (for example, Google DeepMind and Google Brain count as Google). Google's models as CSV.

Google beside the others

All 10 companies →
The same rows for every company. A dash means no source measures it for that company yet: gaps are shown, never filled with a guess. The small number is the rank among the companies in the full table that the source covers.
LensGoogleOpenAIAnthropicxAIMetaMicrosoftDeepSeekAlibabaMistral AINvidia
CapabilityEpoch Capabilities Index, its best model · data166.4#3168.8#1168.8#1156.6#6161.8#4130.5#10155.9#7157.4#5150.0#8146.0#9
OutputNotable AI models released in 2025 · chart15#221#17#45#54#60#104#611#31#93#8
MoneyAnnualized revenue, latest report · data—$50B#2Sep 2026$65B#1Jul 2026$428M#5Sep 2025——$1B#3Sep 2026—$1B#3Sep 2026—
ComputeLargest training run with a published estimate, FLOP · chart5×10²⁵#4Gemini 1.0 Ultra, Dec 20239.4×10²⁶#1GPT-6 Astra, Sep 20263.4×10²⁵#5Claude 3.7 Sonnet, Feb 20255×10²⁶#2Grok 4, Jul 20255.2×10²⁵#3Llama 4 Behemoth (preview), Apr 20258.6×10²³#9Megatron-Turing NLG 530B, Oct 2021 (with Nvidia)9.7×10²⁴#8DeepSeek-V4-Pro, Apr 20261.5×10²⁵#7Qwen3-Max, Sep 20257.7×10²³#10Mixtral 8x7B, Dec 20231.8×10²⁵#6Nemotron-4 340B, Jun 2024
UseUS adults who used its assistant in the past week, Mar 2026 · data20.9%#231.0%#13.0%#64.5%#57.5%#410.5%#30.6%#7———
Safety & policyLowest hallucination rate of its models (lower is better) · chart50%#389%#826%#222%#187%#7—82%#489%#884%#683%#5
More measures · Epoch AI data
Text Arenabest score, Mar 2026 · data1,494#31,481#41,503#11,495#21,335#8—1,424#61,449#51,416#7—
Valuationlatest closed round · data—$852.0B#2$965.0B#1$230.0B#3——$50.0B#4—$24.0B#5—
Equity raisedin all · data—$182.8B#1$139.4B#2$37.0B#3——$7.5B#4—$6.5B#5—
Stafflatest report · data6,000#14,500#23,000#4300#63,400#3—300#6—350#5—
AI chips ownedH100-equivalents · data5.47M#1——0.55M#41.95M#33.32M#2————
Compute availableH100-equivalents, end of 2025 · data1.58M#21.74M#11.19M#30.63M#51.00M#4—————
Weekly use at workemployed US adults, Jul 2026 · data24.2%#233.7%#17.7%#53.3%#69.8%#414.8%#3————
Model record · Epoch AI estimates
Notable models, 2014–2025Epoch AI's count, as in chart 1.1.7194#161#313#77#887#242#47#827#66#1030#5
Frontier modelsmarked frontier in Epoch AI's database33#115#23#53#55#33#50#80#80#84#4
Open weightsshare of its notable models since 202314%8 of 574%2 of 570%0 of 2211%1 of 973%19 of 2633%2 of 6100%10 of 1075%24 of 3267%4 of 683%15 of 18
Largest estimated training cost2023 US dollars, among models Epoch has costed$31M#4Gemini 1.0 Ultra$366M#2GPT-4.5$26M#5Claude 3.5 Sonnet$388M#1Grok 4$53M#3Llama 3.1-405B$4M#9Megatron-Turing NLG 530B$7M#8DeepSeek-R1 (May 2025)$109K#10PLUG$14M#7Mistral Large$21M#6Nemotron-4 340B
Largest training clusterchips for one model, where Epoch records it57,000#3Gemini 1.0 Ultra100,000#2GPT-6 Astra—200,000#1Grok 432,000#4Llama 4 Behemoth (preview)4,480#6Megatron-Turing NLG 530B2,048#7DeepSeek-V3 (Mar 2025)512#8CogView—6,144#5Nemotron-4 340B
Charts on AI Adoptionthat name the company2333228404714

Every chart that names Google

All charts →
23 charts across 5 hubs name Google or one of its models. Each has its source and a CSV of its data.
AI Benchmarks15 charts
  1. 1,503Claude Opus 4.6 tops the Text Arena at 1,503, and the top 15 models span 46 points
  2. 91.2%Gemini-3.1-Pro leads MMLU-Pro at 91.2%, and the top 15 models are within 4.2 points
  3. 93%The best GPQA Diamond score rose from 36% in 2023 to 93% in 2025, above the 81.2% expert baseline
  4. 38.3%The best score on Humanity's Last Exam rose from 8% in 2024 to 38.3% in 2025
  5. 84.6%Gemini 3 Deep Think scored 84.6% on ARC-AGI-2, 7.5 points ahead of the next system
  6. 205GPT-5 solved 205 PlanBench tasks across eight domains, level with the classical planner LAMA's 204
  7. 91.9%Aletheia scored 91.9% on IMO-ProofBench olympiad proofs in early 2026, against 7.1% for GPT-5.1
  8. 64.4%The best agent succeeded on 64.4% of MLE-bench's Kaggle tasks in early 2026, up from 17.1% in 2024
  9. 0.86Arabic.AI's LLM-X topped HELM Arabic with a mean score of 0.86, ahead of Gemini 2.5 Flash at 0.82
  10. 1,314GPT-5.2 leads the Indic LLM Arena with 1,314, ahead of GPT-5.1 (1,298) and Gemini 3 Flash (1,288)
  11. 88.2%Gemini 3.1 Pro Preview scored 88.2% on MMMU, 0.4 points short of the best human expert
  12. 1,288Google's gemini-3-pro leads the Vision Arena at 1,288; Gemini models hold four of the top six places
  13. 22.6%The best of seven AI models read analogue clocks exactly right 22.6% of the time on ClockQA
  14. 69.4%No model reached 70% on MortgageTax in early 2026; Gemini 3.1 Pro Preview led at 69.4%
  15. 87.4%All top 15 models scored above 83% on LegalBench in early 2026; Gemini 3.1 Pro Preview led at 87.4%
AI Models4 charts
  1. 21OpenAI released the most notable AI models in 2025: 21, ahead of Google's 15
  2. 194Google leads with 194 notable AI models since 2014, more than twice Meta's 87
  3. 1,503Anthropic, xAI, Google and OpenAI's best models were within 22 Arena points in March 2026
  4. 42% → 0.3%Google's share of top Hugging Face downloads fell from 42% in 2020 to under 1%
AI Infrastructure2 charts
  1. 8484 notable AI models have been trained on Nvidia A100 chips, more than on any other accelerator
  2. 1.21 WhA day of eight medium GPT-4o queries uses about as much energy as charging two phones
AI Research1 chart
  1. 38.8%Gemini 2.5 Pro scored 38.8% on PaperArena as a multi-agent system, the best of nine models tested
AI in Healthcare1 chart
  1. 92%Google showed an AI Overview on 92% of symptom and common health searches analysed

The other companies, same lens

AI Companies →
OpenAI168.8Epoch Capabilities Index65 ledger entriesOpen tracker →Anthropic168.8Epoch Capabilities Index61 ledger entriesOpen tracker →xAI156.6Epoch Capabilities Index37 ledger entriesOpen tracker →Meta161.8Epoch Capabilities Index33 ledger entriesOpen tracker →Microsoft130.5Epoch Capabilities Index4 ledger entriesOpen tracker →DeepSeek155.9Epoch Capabilities Index35 ledger entriesOpen tracker →Alibaba157.4Epoch Capabilities Index33 ledger entriesOpen tracker →Mistral AI150.0Epoch Capabilities Index26 ledger entriesOpen tracker →Nvidia146.0Epoch Capabilities Index7 ledger entriesOpen tracker →

How the ledger works

  1. One lens for everyone. Six lenses, same order, same definitions, for every company. No company gets an extra lens for a product it wants to talk about.
  2. Changes, not news. An entry must move a number or a rank on a lens. Launch events, interviews and hires do not count unless they do.
  3. Evidence tier on every line. Independent measure, independent estimate or company data. A company's own claims are replaced when an independent score arrives.
  4. Gaps stay visible. If no source measures a lens for a company, the page says so.
Disclosure: AI Adoption uses AI tools in its research and editorial work, including products from some of the companies tracked here. No company has any say in its page or its ledger, and the same rules apply to every company. Editorial standards · Data sources
Copied with its source. Want the next number like this in your inbox?Get Friday's numbers