Skip to Sidebar Skip to Content
Evidence · Guidance · Learning

Know what works in AI.

Sourced evidence and practical guidance for the people putting AI to work.

Every number traced to its sourceNo one pays to be ratedChecked 2 Nov 2026
AI by CountryWho leads, top 5

The US and China lead on Progress at 77 and 74 out of 100; Germany and France lead the rulebook at 94.

Open AI by Country →
The AI RecordLive

News with memory: each event we report, tied to what came before it and to the numbers it changes.

Thread · updatedAI Layoffs: Job Cuts Companies Link to AIThread · updatedAI Copyright CasesThread · updatedAI Chatbot Rules for ChildrenOpen the Record →
AI CompaniesLedgers, 2023–2026

On the Epoch Capabilities Index the best models from OpenAI, Anthropic and Google sit within 3.0 points. Anthropic's annualized revenue reached $65B (Jul 2026), OpenAI's $50B (Sep 2026).

IndependentCompany dataEach dot is a month with changes; bigger dot, more changes.
OpenAIAnthropicGooglexAIMetaMicrosoftDeepSeekAlibabaMistral AINvidiaOpenAI · Jan 2023: Had about 375 staffOpenAI · Mar 2023: Said it had 50M weekly active users of ChatGPT | 15 Mar 2023: GPT-4 (Mar 2023) released, trained with an estimated 2.1×10²⁵ FLOPOpenAI · 26 Apr 2023: Closed a funding round at a $27.0B post-money valuation, raising $300M in equityOpenAI · May 2023: Took first place on the Text Arena with 1,322 points, its first month in this seriesOpenAI · 13 Jun 2023: GPT-4 (Jun 2023) released, marked a frontier model in Epoch AI's database, trained with an estimated 2.1×10²⁵ FLOPOpenAI · 29 Aug 2023: Annualized revenue reached $1B, from $200M on 1 Mar 2023OpenAI · Sep 2023: Said it had 80M weekly active users of ChatGPTOpenAI · 10 Oct 2023: Annualized revenue reached $1.3B, from $1B on 29 Aug 2023OpenAI · Nov 2023: Had about 800 staff | 6 Nov 2023: GPT-4 Turbo (Nov 2023) released, marked a frontier model in Epoch AI's databaseOpenAI · 31 Dec 2023: Annualized revenue reached $2B, from $1.6B on 30 Dec 2023 | 30 Dec 2023: Annualized revenue reached $1.6B, from $1.3B on 10 Oct 2023OpenAI · 16 Feb 2024: Closed a funding round at a $80.0B post-money valuation | Feb 2024: Said it processed 133B tokens a dayOpenAI · Mar 2024: Said it had 100M monthly active users | Mar 2024: Said it had 125M weekly active users of ChatGPT | Mar 2024: Fell to #2 of 7 on the Text Arena, 8 points behind Anthropic3OpenAI · 9 Apr 2024: GPT-4 Turbo (Apr 2024) released, marked a frontier model in Epoch AI's database | Apr 2024: Took first place on the Text Arena with 1,338 pointsOpenAI · 13 May 2024: GPT-4o released, marked a frontier model in Epoch AI's databaseOpenAI · Jun 2024: Said it had 350M monthly active users | Jun 2024: Said it processed 451M messages a day on ChatGPT | 12 Jun 2024: Annualized revenue reached $3.4B, from $2B on 31 Dec 20233OpenAI · Jul 2024: Had about 1,500 staffOpenAI · 15 Aug 2024: Annualized revenue reached $3.6B, from $3.4B on 12 Jun 2024 | 6 Aug 2024: GPT-4o (Aug 2024) released, marked a frontier model in Epoch AI's databaseOpenAI · Sep 2024: Had about 1,700 staff | Sep 2024: Said it had 240M weekly active users of ChatGPT | 12 Sep 2024: Annualized revenue reached $4B, from $3.6B on 15 Aug 20243OpenAI · 2 Oct 2024: Closed a funding round at a $157.0B post-money valuation, raising $6.6B in equityOpenAI · 20 Nov 2024: GPT-4o (Nov 2024) released, marked a frontier model in Epoch AI's databaseOpenAI · 31 Dec 2024: Annualized revenue reached $5.5B, from $4B on 12 Sep 2024 | 2024: Reported compute spend was $5.8B in 2024 | Dec 2024: Said it had 300M weekly active users of ChatGPT3OpenAI · 29 Jan 2025: GPT-4o (Jan 2025) released, marked a frontier model in Epoch AI's databaseOpenAI · 27 Feb 2025: GPT-4.5 released, trained with an estimated 3.8×10²⁶ FLOPOpenAI · 31 Mar 2025: Closed a funding round at a $300.0B post-money valuation, raising $10.0B in equity | Mar 2025: Said it had 160M daily active users of ChatGPT | 27 Mar 2025: GPT-4o (Mar 2025) released, marked a frontier model in Epoch AI's database | Mar 2025: Said it had 430M weekly active users of ChatGPT4OpenAI · May 2025: Fell to #2 of 8 on the Text Arena, 34 points behind GoogleOpenAI · 2025 study: GPT-5 solved 205 PlanBench tasks, level with the classical planner LAMA's 204 | 2025 study: GPT-4o had the best safe success rate on ResponsibleRobotBench, 0.64 | 2025: One medium-length GPT-4o query uses about 1.21 Wh of energy | 2025: GPT-4o's inference uses an estimated 1.33 to 1.58 million kilolitres of water a year | Jun 2025: Said it processed 2.63B messages a day on ChatGPT | 9 Jun 2025: Annualized revenue reached $10B, from $5.5B on 31 Dec 20246OpenAI · 30 Jul 2025: Annualized revenue reached $12B, from $10B on 9 Jun 2025 | Jul 2025: Said it had 770M weekly active users of ChatGPT | Jul 2025: Had about 3,000 staff | Jul 2025: Took first place on the Text Arena with 1,482 points4OpenAI · 7 Aug 2025: GPT-5 released, trained with an estimated 6.6×10²⁵ FLOP | 5 Aug 2025: gpt-oss-120b released, trained with an estimated 4.9×10²⁴ FLOP | Aug 2025: Said it had 700M weekly active users of ChatGPT | Aug 2025: Said it processed 3B messages a day on ChatGPT | Aug 2025: Fell to #3 of 8 on the Text Arena, 12 points behind Google | 1 Aug 2025: Annualized revenue reached $13B, from $12B on 30 Jul 2025 | 1 Aug 2025: Closed a funding round at a $300.0B post-money valuation, raising $8.3B in equity7OpenAI · Oct 2025: Said it had 800M weekly active users of ChatGPT | Oct 2025: Said it processed 8.6T tokens a day on API | 1 Oct 2025: Closed a funding round at a $500.0B post-money valuation3OpenAI · Nov 2025: Fell to #4 of 8 on the Text Arena, 31 points behind GoogleOpenAI · 31 Dec 2025: Annualized revenue reached $21.4B, from $19B on 18 Dec 2025 | End of 2025: Had compute equal to about 1.74 million H100s available at the end of 2025 | 2025: Reported compute spend was $16.3B in 2025, up from $5.8B | Dec 2025: Had about 4,000 staff | 2025: Released 22 notable AI models in 2025, the most of any organisation | 2025: GPT-5.2 Pro solved 31.3% of FrontierMath Tier 4, the hardest maths tier, the best result of 2025 | 2025: Agentic AI overtook ChatGPT as the most-cited AI agent skill in US AI job postings | 2025: 14,376 US AI job postings named ChatGPT as a skill in 2025 | 18 Dec 2025: Annualized revenue reached $19B, from $13B on 1 Aug 2025 | Dec 2025: GPT-5.2 (high) scored 97% on MathArena's fresh competition problems | Dec 2025: Said it had 900M weekly active users of ChatGPT11OpenAI · Early 2026: GPT 5.1 led CaseLaw v2 legal research at 73.4%OpenAI · 28 Feb 2026: Annualized revenue reached $25B, from $21.4B on 31 Dec 2025 | Feb 2026: Said it had 920M weekly active users of ChatGPT | 27 Feb 2026: Closed a funding round at a $840.0B post-money valuation, raising $110.0B in equity3OpenAI · 31 Mar 2026: Closed a funding round at a $852.0B post-money valuation, raising $12.0B in equity | Mar 2026: Had about 4,500 staff | Mar 2026: GPT-5.4 High read 50.6% of ClockBench clocks correctly, the best of 15 models | Mar 2026: 31.0% of US adults said they had used ChatGPT in the past week | 2026 test: Its most reliable model on AA-Omniscience, GPT-5.4 (xhigh), had a 89% hallucination rate | Mar 2026: Still #4 of 8 with 1,481; the top four labs sat within 22 Arena points6OpenAI · Jul 2026: Said it had 1B weekly active users of ChatGPT | Jul 2026: 33.7% of employed US adults said they had used ChatGPT in the past week | 15 Jul 2026: Annualized revenue reached $30B, from $25B on 28 Feb 20263OpenAI · 30 Sep 2026: Annualized revenue reached $50B, from $30B on 15 Jul 2026 | Sep 2026: Said it had 1.2B weekly active users of ChatGPT | 3 Sep 2026: GPT-6 Astra scored 169.1 on the Epoch Capabilities Index, its best model | 3 Sep 2026: GPT-6 Astra released, trained with an estimated 9.4×10²⁶ FLOP | 3 Sep 2026: Training GPT-6 Astra drew an estimated 233 MW of power5OpenAI · 24 Oct 2026: Listed 796 open roles on its careers siteAnthropic · Jan 2023: Had about 85 staffAnthropic · 23 May 2023: Closed a funding round at a $4.5B post-money valuation, raising $450M in equityAnthropic · 11 Jul 2023: Claude 2 released, trained with an estimated 3.9×10²⁴ FLOPAnthropic · Oct 2023: Had about 192 staffAnthropic · 2023: Reported compute spend was $285M in 2023 | Dec 2023: Had about 300 staffAnthropic · 11 Jan 2024: Closed a funding round at a $18.4B post-money valuation, raising $750M in equity | 1 Jan 2024: Annualized revenue was estimated at $87M, against $100M on 1 Oct 2023Anthropic · 4 Mar 2024: Claude 3 Opus released, marked a frontier model in Epoch AI's database | Mar 2024: Took first place on the Text Arena with 1,333 points, its first month in this seriesAnthropic · Apr 2024: Fell to #2 of 7 on the Text Arena, 15 points behind OpenAIAnthropic · May 2024: Had about 500 staffAnthropic · 20 Jun 2024: Claude 3.5 Sonnet released, trained with an estimated 2.7×10²⁵ FLOPAnthropic · Sep 2024: Fell to #3 of 8 on the Text Arena, 50 points behind OpenAIAnthropic · Oct 2024: Rose to #2 of 8 on the Text Arena, 20 points behind OpenAIAnthropic · 31 Dec 2024: Annualized revenue reached $1B, from $87M on 1 Jan 2024 | Dec 2024: Had about 915 staff | 2024: Reported compute spend was $2.5B in 2024, up from $285M3Anthropic · Jan 2025: Fell to #3 of 8 on the Text Arena, 40 points behind OpenAI | Jan 2025: First reading: automation was 41% of sampled Claude.ai conversations, augmentation 55%Anthropic · 24 Feb 2025: Claude 3.7 Sonnet released, trained with an estimated 3.4×10²⁵ FLOPAnthropic · 31 Mar 2025: Annualized revenue reached $2B, from $1.4B on 1 Mar 2025 | 3 Mar 2025: Closed a funding round at a $61.5B post-money valuation, raising $3.5B in equity | 1 Mar 2025: Annualized revenue reached $1.4B, from $1B on 31 Dec 2024 | Mar 2025: Automation was 42% of sampled conversations, augmentation 55%4Anthropic · Apr 2025: Fell to #4 of 8 on the Text Arena, 70 points behind OpenAIAnthropic · 30 May 2025: Annualized revenue reached $3B, from $2B on 31 Mar 2025 | May 2025: Had about 1,300 staff | May 2025: Rose to #3 of 8 on the Text Arena, 69 points behind Google3Anthropic · Jun 2025: Fell to #4 of 8 on the Text Arena, 52 points behind GoogleAnthropic · 29 Jul 2025: Annualized revenue reached $5B, from $4B on 1 Jul 2025 | Jul 2025: Fell to #5 of 8 on the Text Arena, 59 points behind OpenAI | 1 Jul 2025: Annualized revenue reached $4B, from $3B on 30 May 20253Anthropic · Aug 2025: Rose to #2 of 8 on the Text Arena, 3 points behind Google | Aug 2025: Automation overtook augmentation, 49% to 47%Anthropic · 2 Sep 2025: Closed a funding round at a $183.0B post-money valuation, raising $13.0B in equityAnthropic · 21 Oct 2025: Annualized revenue reached $7B, from $5B on 29 Jul 2025Anthropic · Nov 2025: Fell to #3 of 8 on the Text Arena, 26 points behind Google | Nov 2025: Augmentation moved back ahead, 52% to 45% | Nov 2025: Education tasks reached 15% of sampled Claude conversations3Anthropic · 31 Dec 2025: Annualized revenue reached $9B, from $7B on 21 Oct 2025 | End of 2025: Had compute equal to about 1.19 million H100s available at the end of 2025 | 2025: Reported compute spend was $6.8B in 2025, up from $2.5B | 2025: Released 7 notable AI models in 2025, #4 by count | Dec 2025: Had about 2,300 staff5Anthropic · Early 2026: Claude Opus 4.6 solved 93% of Cybench security challenges unguided | Early 2026: Claude Opus 4.6 led Vibe Code Bench at 57.6% | Early 2026: Claude Opus 4.5 led τ-bench tool use at 70.2%, 0.3 points ahead of GPT-5.2 | Early 2026: Claude-Opus-4.5 led the Berkeley Function Calling Leaderboard at 77.5% | Early 2026: Claude Sonnet 4.6 led Finance Agent at 63.3% | Early 2026: Claude Sonnet 4.6 led TaxEval v2 at 77.1%6Anthropic · 2026: Claude Sonnet 4.6 set the best OSWorld score in Epoch AI's data, 72.1% | Feb 2026: Claude 4.5 Opus led SWE-bench Verified at 76.8% | 12 Feb 2026: Annualized revenue reached $14B, from $9B on 31 Dec 2025 | 12 Feb 2026: Closed a funding round at a $380.0B post-money valuation, raising $30.0B in equity | Feb 2026: Took first place on the Text Arena with 1,504 points5Anthropic · Mar 2026: 3.0% of US adults said they had used Claude in the past week | 2026 test: Its most reliable model on AA-Omniscience, Claude 4.5 Haiku, had a 26% hallucination rate | 3 Mar 2026: Annualized revenue reached $19B, from $14B on 12 Feb 2026 | Mar 2026: Still first with 1,503; the top four labs sat within 22 Arena points4Anthropic · 24 Apr 2026: Closed a funding round at a $380.0B post-money valuation, raising $10.0B in equity | 20 Apr 2026: Closed a funding round at a $380.0B post-money valuation, raising $5.0B in equity | 6 Apr 2026: Annualized revenue reached $30B, from $19B on 3 Mar 20263Anthropic · 28 May 2026: Closed a funding round at a $965.0B post-money valuation, raising $65.0B in equity | 15 May 2026: Annualized revenue reached $47B, from $30B on 6 Apr 2026Anthropic · 31 Jul 2026: Annualized revenue reached $65B, from $47B on 15 May 2026 | Jul 2026: 7.7% of employed US adults said they had used Claude in the past weekAnthropic · Sep 2026: Had about 5,300 staff | 22 Sep 2026: Claude Opus 5.5 scored 168.4 on the Epoch Capabilities Index, its best modelAnthropic · 24 Oct 2026: Listed 681 open roles on its careers siteGoogle · 10 May 2023: PaLM 2 released, trained with an estimated 7.3×10²⁴ FLOP | May 2023: Placed #2 of 3 on the Text Arena with 1,117 points, its first month in this seriesGoogle · Aug 2023: Fell to #3 of 3 on the Text Arena, 197 points behind OpenAIGoogle · Oct 2023: Fell to #4 of 5 on the Text Arena, 184 points behind OpenAIGoogle · 14 Nov 2023: GraphCast released, trained with an estimated 2.1×10²² FLOPGoogle · 14 Dec 2023: FunSearch released, trained with an estimated 3.9×10²³ FLOP | 6 Dec 2023: Gemini 1.0 Ultra released, trained with an estimated 5×10²⁵ FLOP | 6 Dec 2023: Training Gemini 1.0 Ultra drew an estimated 38 MW of power | Dec 2023: Rose to #2 of 5 on the Text Arena, 95 points behind OpenAI4Google · Jan 2024: Fell to #3 of 6 on the Text Arena, 117 points behind OpenAIGoogle · 15 Feb 2024: Gemini 1.5 Pro released, marked a frontier model in Epoch AI's database | Feb 2024: Fell to #4 of 6 on the Text Arena, 113 points behind OpenAIGoogle · Mar 2024: Fell to #5 of 7 on the Text Arena, 113 points behind AnthropicGoogle · Apr 2024: Said it processed 323B tokens a day on Google (all AI products) | Apr 2024: Fell to #6 of 7 on the Text Arena, 109 points behind OpenAIGoogle · May 2024: Had about 2,600 staff at Google DeepMind | 8 May 2024: AlphaFold 3 released, trained with an estimated 4.1×10²² FLOP | May 2024: Rose to #3 of 7 on the Text Arena, 23 points behind OpenAI | 1 May 2024: GenCast released, trained with an estimated 8.2×10²⁰ FLOP4Google · Aug 2024: Fell to #5 of 8 on the Text Arena, 13 points behind OpenAIGoogle · Sep 2024: Rose to #2 of 8 on the Text Arena, 35 points behind OpenAIGoogle · Oct 2024: Fell to #3 of 8 on the Text Arena, 31 points behind OpenAIGoogle · Dec 2024: Fell to #4 of 8 on the Text Arena, 56 points behind OpenAIGoogle · Mar 2025: Said it had 35M daily active users of Gemini | Mar 2025: Had about 5,600 staff at Google DeepMind | Mar 2025: Said it had 350M monthly active users of Gemini | Mar 2025: Fell to #5 of 8 on the Text Arena, 71 points behind OpenAI4Google · Apr 2025: Said it processed 16T tokens a day on Google (all AI products) | Apr 2025: Fell to #6 of 8 on the Text Arena, 92 points behind OpenAIGoogle · May 2025: Said it had 400M monthly active users of Gemini | May 2025: Took first place on the Text Arena with 1,492 pointsGoogle · 2025 study: Gemini 2.5 Pro scored 38.8% on PaperArena as a multi-agent system, the best of nine models | 2025 study: Gemini-2.0 read analogue clocks exactly right 22.6% of the time, the best of seven models | 2025 study: AI Overviews appeared on 84% to 92% of the health-related Google searches analysed3Google · Jul 2025: Said it processed 33T tokens a day on Google (all AI products) | Jul 2025: Said it had 450M monthly active users of Gemini | Jul 2025: Fell to #2 of 8 on the Text Arena, 20 points behind OpenAI3Google · Aug 2025: Had about 6,000 staff at Google DeepMind | Aug 2025: Took first place on the Text Arena with 1,461 pointsGoogle · Sep 2025: Said it processed 43T tokens a day on Google (all AI products) | Q3 2025: Google's share of downloads of the top Hugging Face models fell to 0.3%Google · Oct 2025: Said it had 650M monthly active users of GeminiGoogle · End of 2025: Had compute equal to about 1.58 million H100s available at the end of 2025 | 2014–2025: Google has released 194 notable AI models since 2014, the most of any organisation | 2025: Released 15 notable AI models in 2025, #2 by count | 2025: Gemini 3 Pro set the best score on Humanity's Last Exam, 38.3%4Google · Early 2026: Gemini-3.1-Pro led MMLU-Pro at 91.2% | Early 2026: Gemini 3.1 Pro Preview led LegalBench at 87.4% | Early 2026: Gemini 3.1 Pro Preview led MortgageTax at 69.4%; no model reached 70%3Google · Feb 2026: Gemini 3 Deep Think scored 84.6% on ARC-AGI-2, 7.5 points ahead of the next system | Feb 2026: Gemini 3.1 Pro Preview scored 88.2% on MMMU, 0.4 points short of the best human expert | Feb 2026: gemini-3-pro led the Vision Arena at 1,288; Gemini models held four of the top six places | Feb 2026: Fell to #2 of 8 on the Text Arena, 4 points behind Anthropic4Google · Mar 2026: Owned AI chips equal to about 5.47 million Nvidia H100s | Mar 2026: 20.9% of US adults said they had used Google Gemini in the past week | 2026 test: Its most reliable model on AA-Omniscience, Gemini 3.1 Pro Preview, had a 50% hallucination rate | Mar 2026: Fell to #3 of 8 on the Text Arena, 9 points behind Anthropic4Google · 2026: Google DeepMind's unreleased AI co-mathematician solved 47.9% of FrontierMath Tier 4, the hardest maths tierGoogle · Jul 2026: 24.2% of employed US adults said they had used Google Gemini in the past weekGoogle · 30 Sep 2026: Gemini 4 Argon scored 166.1 on the Epoch Capabilities Index, its best model | 2026: Gemini 4 Argon (high) set the best GPQA Diamond score, 96.4%, 15.2 points above expert validatorsxAI · Jul 2023: Had about 12 staffxAI · 4 Nov 2023: Grok-1 released, trained with an estimated 2.9×10²⁴ FLOPxAI · 27 May 2024: Closed a funding round at a $24.0B post-money valuation, raising $6.0B in equityxAI · Jun 2024: Had about 15 staffxAI · 13 Aug 2024: Grok-2 released, trained with an estimated 3×10²⁵ FLOP | Aug 2024: Placed #4 of 8 on the Text Arena with 1,331 points, its first month in this seriesxAI · Sep 2024: Fell to #5 of 8 on the Text Arena, 56 points behind OpenAIxAI · 20 Nov 2024: Closed a funding round at a $45.0B post-money valuation, raising $6.0B in equityxAI · Dec 2024: Fell to #6 of 8 on the Text Arena, 78 points behind OpenAIxAI · 17 Feb 2025: Grok 3 released, trained with an estimated 3.5×10²⁶ FLOP | 17 Feb 2025: Training Grok 3 drew an estimated 110 MW of power | Feb 2025: Fell to #7 of 8 on the Text Arena, 74 points behind OpenAI3xAI · 31 Mar 2025: Annualized revenue reached $172M, from $100M on 20 Nov 2024 | 28 Mar 2025: Closed a funding round at a $113.0B post-money valuation | Mar 2025: Said it had 15M daily active users of Grok | Mar 2025: Had about 300 staff | Mar 2025: Rose to #2 of 8 on the Text Arena, 30 points behind OpenAI5xAI · May 2025: Fell to #4 of 8 on the Text Arena, 80 points behind GooglexAI · 30 Jun 2025: Annualized revenue reached $236M, from $172M on 31 Mar 2025 | 30 Jun 2025: Closed a funding round at a $150.0B post-money valuation, raising $5.0B in equity | Jun 2025: Fell to #5 of 8 on the Text Arena, 58 points behind Google3xAI · 2025: Training Grok 4 emitted an estimated 72,816 tonnes of CO₂ equivalent | 9 Jul 2025: Grok 4 released, trained with an estimated 5×10²⁶ FLOP | Jul 2025: Rose to #4 of 8 on the Text Arena, 50 points behind OpenAI3xAI · Aug 2025: Fell to #6 of 8 on the Text Arena, 34 points behind GooglexAI · 30 Sep 2025: Annualized revenue reached $428M, from $236M on 30 Jun 2025 | Sep 2025: Said it had 64M monthly active users of Grok | Sep 2025: Rose to #5 of 8 on the Text Arena, 28 points behind Google3xAI · Oct 2025: Fell to #6 of 8 on the Text Arena, 32 points behind GooglexAI · Nov 2025: Grok 4.1 Fast (Reasoning) was the best model released in November 2025 on MathArena, at 86.4% | Nov 2025: Rose to #2 of 8 on the Text Arena, 10 points behind GooglexAI · Dec 2025: Owned AI chips equal to about 0.55 million Nvidia H100s | End of 2025: Had compute equal to about 0.63 million H100s available at the end of 2025 | 2025: Released 5 notable AI models in 2025, #5 by count | 2025: Grok 3 and Grok 4 were the largest notable models of 2025, at an estimated 3 trillion parameters4xAI · 6 Jan 2026: Closed a funding round at a $230.0B post-money valuation, raising $20.0B in equityxAI · Feb 2026: Fell to #3 of 8 on the Text Arena, 11 points behind AnthropicxAI · Mar 2026: 4.5% of US adults said they had used Grok in the past week | 2026 test: Its most reliable model on AA-Omniscience, Grok 4.20 Beta 0305, had a 22% hallucination rate | Mar 2026: Rose to #2 of 8 on the Text Arena, 8 points behind Anthropic3xAI · Jul 2026: 3.3% of employed US adults said they had used Grok in the past weekxAI · 12 Aug 2026: Grok 4.6 scored 156.9 on the Epoch Capabilities Index, its best modelMeta · 24 Feb 2023: LLaMA-65B released, trained with an estimated 5.5×10²³ FLOPMeta · May 2023: Placed #3 of 3 on the Text Arena with 969 points, its first month in this seriesMeta · 18 Jul 2023: Llama 2-70B released, trained with an estimated 8.1×10²³ FLOPMeta · Aug 2023: Rose to #2 of 3 on the Text Arena, 146 points behind OpenAIMeta · Oct 2023: Fell to #3 of 5 on the Text Arena, 166 points behind OpenAIMeta · Dec 2023: Fell to #5 of 5 on the Text Arena, 166 points behind OpenAIMeta · Feb 2024: Fell to #6 of 6 on the Text Arena, 175 points behind OpenAIMeta · Mar 2024: Fell to #7 of 7 on the Text Arena, 181 points behind AnthropicMeta · 18 Apr 2024: Llama 3-70B released, trained with an estimated 7.9×10²⁴ FLOP | Apr 2024: Rose to #3 of 7 on the Text Arena, 59 points behind OpenAIMeta · May 2024: Fell to #4 of 7 on the Text Arena, 70 points behind OpenAIMeta · 23 Jul 2024: Llama 3.1-405B released, trained with an estimated 3.8×10²⁵ FLOP | Jul 2024: Rose to #3 of 7 on the Text Arena, 14 points behind OpenAIMeta · Sep 2024: Said it had 400M monthly active users of Meta AI assistant | Sep 2024: Said it had 185M weekly active users of Meta AI assistant | Sep 2024: Fell to #4 of 8 on the Text Arena, 53 points behind OpenAI3Meta · Oct 2024: Said it had 500M monthly active users of Meta AI assistantMeta · 2024: Training Llama 3.1 405B emitted an estimated 8,930 tonnes of CO₂ | Dec 2024: Said it had 600M monthly active users of Meta AI assistant | 6 Dec 2024: Llama 3.3 70B released, trained with an estimated 6.9×10²⁴ FLOP | Dec 2024: Fell to #5 of 8 on the Text Arena, 73 points behind OpenAI4Meta · Jan 2025: Said it had 700M monthly active users of Meta AI assistantMeta · Feb 2025: Fell to #6 of 8 on the Text Arena, 71 points behind OpenAIMeta · Mar 2025: Said it had 100M daily active users of Meta AI assistant | Mar 2025: Fell to #7 of 8 on the Text Arena, 109 points behind OpenAIMeta · 5 Apr 2025: Llama 4 Behemoth (preview) released, trained with an estimated 5.2×10²⁵ FLOP | 5 Apr 2025: Llama 4 Maverick released, trained with an estimated 2.2×10²⁴ FLOP | 5 Apr 2025: Training Llama 4 Behemoth (preview) drew an estimated 44 MW of power | Apr 2025: Had about 1,000 staff at FAIR4Meta · May 2025: Said it had 1B monthly active users of Meta AI assistant | May 2025: Fell to #8 of 8 on the Text Arena, 159 points behind GoogleMeta · 2025 study: Llama 3.2-Vision answered 11.7% of CalendarQA date questions correctlyMeta · Jul 2025: Had about 3,400 staffMeta · Oct 2025: Said it had 1B monthly active users of Meta AI assistantMeta · End of 2025: Had compute equal to about 1.00 million H100s available at the end of 2025 | 2014–2025: Meta has released 87 notable AI models since 2014, second only to Google's 194 | 2025: Released 4 notable AI models in 2025, #6 by count3Meta · Mar 2026: Owned AI chips equal to about 1.95 million Nvidia H100s | Mar 2026: 7.5% of US adults said they had used Meta AI in the past week | 2026 test: Its most reliable model on AA-Omniscience, Llama 4 Maverick, had a 87% hallucination rate | Mar 2026: Still #8 of 8 with 1,335; the top four labs sat within 22 Arena points4Meta · Jul 2026: 9.8% of employed US adults said they had used Meta AI in the past weekMeta · 12 Oct 2026: Muse Wonder scored 161.7 on the Epoch Capabilities Index, its best modelMicrosoft · 12 Dec 2024: Phi-4 scored 130.6 on the Epoch Capabilities Index, its best modelMicrosoft · Mar 2026: Owned AI chips equal to about 3.32 million Nvidia H100s | Mar 2026: 10.5% of US adults said they had used Microsoft Copilot in the past weekMicrosoft · Jul 2026: 14.8% of employed US adults said they had used Microsoft Copilot in the past weekDeepSeek · Jan 2024: Placed #4 of 6 on the Text Arena with 1,189 points, its first month in this seriesDeepSeek · Feb 2024: Fell to #5 of 6 on the Text Arena, 143 points behind OpenAIDeepSeek · Mar 2024: Fell to #6 of 7 on the Text Arena, 148 points behind AnthropicDeepSeek · Apr 2024: Fell to #7 of 7 on the Text Arena, 153 points behind OpenAIDeepSeek · 17 Jun 2024: DeepSeek-Coder-V2 236B released, trained with an estimated 1.3×10²⁴ FLOP | Jun 2024: Rose to #6 of 7 on the Text Arena, 83 points behind OpenAIDeepSeek · Jul 2024: Fell to #7 of 7 on the Text Arena, 82 points behind OpenAIDeepSeek · Aug 2024: Fell to #8 of 8 on the Text Arena, 82 points behind OpenAIDeepSeek · 6 Sep 2024: DeepSeek-V2.5 released, trained with an estimated 1.8×10²⁴ FLOPDeepSeek · 2024: Training DeepSeek v3 emitted an estimated 597 tonnes of CO₂, a fifteenth of Llama 3.1 405B's 8,930 | 24 Dec 2024: DeepSeek-V3 released, trained with an estimated 3.3×10²⁴ FLOP | Dec 2024: Rose to #3 of 8 on the Text Arena, 47 points behind OpenAI3DeepSeek · 20 Jan 2025: DeepSeek-R1 released, trained with an estimated 3.5×10²⁴ FLOP | Jan 2025: Rose to #2 of 8 on the Text Arena, 8 points behind OpenAIDeepSeek · Feb 2025: Said it processed 788B tokens a day on Chat and API for V3 and R1DeepSeek · Mar 2025: Said it had 10M daily active users of DeepSeek | Mar 2025: Had about 160 staff | Mar 2025: Fell to #4 of 8 on the Text Arena, 46 points behind OpenAI3DeepSeek · Apr 2025: Rose to #3 of 8 on the Text Arena, 57 points behind OpenAIDeepSeek · May 2025: Fell to #5 of 8 on the Text Arena, 91 points behind GoogleDeepSeek · 2025 study: DeepSeek R1 solved 157 PlanBench tasks, against 205 for GPT-5 and 204 for the classical planner LAMA | 2025 study: Disguising PlanBench tasks cut DeepSeek R1's solved total from 157 to 93 | Jun 2025: Rose to #3 of 8 on the Text Arena, 49 points behind Google3DeepSeek · Jul 2025: Fell to #6 of 8 on the Text Arena, 65 points behind OpenAIDeepSeek · Aug 2025: Rose to #4 of 8 on the Text Arena, 30 points behind GoogleDeepSeek · Sep 2025: Fell to #6 of 8 on the Text Arena, 35 points behind GoogleDeepSeek · Oct 2025: Rose to #5 of 8 on the Text Arena, 31 points behind GoogleDeepSeek · Nov 2025: Fell to #6 of 8 on the Text Arena, 68 points behind GoogleDeepSeek · 2025: Released 4 notable AI models in 2025, #6 by count | 2025: DeepSeek V3.2 Exp used 23.2 Wh per medium-length prompt, the most of the 15 most energy-intensive 2025 modelsDeepSeek · Mar 2026: 0.6% of US adults said they had used DeepSeek in the past week | 2026 test: Its most reliable model on AA-Omniscience, DeepSeek V3.2, had a 82% hallucination rate | Mar 2026: Still #6 of 8 with 1,424; the top four labs sat within 22 Arena points3DeepSeek · 24 Apr 2026: DeepSeek-V4-Pro released, trained with an estimated 9.7×10²⁴ FLOPDeepSeek · Jun 2026: Had about 300 staff | 16 Jun 2026: Closed a funding round at a $50.0B post-money valuation, raising $7.5B in equityDeepSeek · 24 Sep 2026: Annualized revenue reached $1B, from $450M on 14 Jul 2026 | 9 Sep 2026: DeepSeek V4.1 Flash released, trained with an estimated 4.3×10²⁴ FLOPDeepSeek · 22 Oct 2026: Closed a funding round at a $90.0B post-money valuation, raising $15.9B in equity | 17 Oct 2026: DeepSeek V4.1 Flash 1017 scored 158.8 on the Epoch Capabilities Index, its best modelAlibaba · Oct 2023: Placed #2 of 5 on the Text Arena with 1,154 points, its first month in this seriesAlibaba · 30 Nov 2023: Qwen-72B released, trained with an estimated 1.3×10²⁴ FLOPAlibaba · Dec 2023: Fell to #4 of 5 on the Text Arena, 163 points behind OpenAIAlibaba · Jan 2024: Fell to #6 of 6 on the Text Arena, 193 points behind OpenAIAlibaba · Feb 2024: Rose to #3 of 6 on the Text Arena, 95 points behind OpenAIAlibaba · Mar 2024: Fell to #4 of 7 on the Text Arena, 95 points behind AnthropicAlibaba · Apr 2024: Fell to #5 of 7 on the Text Arena, 103 points behind OpenAIAlibaba · May 2024: Fell to #6 of 7 on the Text Arena, 103 points behind OpenAIAlibaba · 7 Jun 2024: Qwen2-72B released, trained with an estimated 3×10²⁴ FLOP | Jun 2024: Rose to #5 of 7 on the Text Arena, 81 points behind OpenAIAlibaba · Jul 2024: Fell to #6 of 7 on the Text Arena, 79 points behind OpenAIAlibaba · Aug 2024: Fell to #7 of 8 on the Text Arena, 80 points behind OpenAIAlibaba · 19 Sep 2024: Qwen2.5-72B released, trained with an estimated 7.8×10²⁴ FLOPAlibaba · Oct 2024: Rose to #6 of 8 on the Text Arena, 74 points behind OpenAIAlibaba · Dec 2024: Fell to #7 of 8 on the Text Arena, 93 points behind OpenAIAlibaba · Feb 2025: Rose to #5 of 8 on the Text Arena, 41 points behind OpenAIAlibaba · Mar 2025: Fell to #6 of 8 on the Text Arena, 72 points behind OpenAIAlibaba · 29 Apr 2025: Qwen3-235B-A22B released, trained with an estimated 4.8×10²⁴ FLOP | Apr 2025: Rose to #5 of 8 on the Text Arena, 91 points behind OpenAIAlibaba · May 2025: Fell to #7 of 8 on the Text Arena, 120 points behind GoogleAlibaba · 2025 study: Qwen-72B was the best open-source model on ResponsibleRobotBench, at 0.35 safe success | 2025 study: Qwen2-VL-7B read none of 62 analogue clocks exactly right on ClockQA | Jun 2025: Rose to #6 of 8 on the Text Arena, 73 points behind Google3Alibaba · Jul 2025: Rose to #3 of 8 on the Text Arena, 44 points behind OpenAIAlibaba · Aug 2025: Fell to #5 of 8 on the Text Arena, 31 points behind GoogleAlibaba · 5 Sep 2025: Qwen3-Max released, trained with an estimated 1.5×10²⁵ FLOP | Sep 2025: Rose to #4 of 8 on the Text Arena, 17 points behind GoogleAlibaba · Nov 2025: Fell to #5 of 8 on the Text Arena, 59 points behind GoogleAlibaba · 2025: Released 12 notable AI models in 2025, #3 by countAlibaba · Mar 2026: Qwen 3-VL 235B Instruct read 39.4% of ClockBench clocks correctly, second of 15 models | 2026 test: Its most reliable model on AA-Omniscience, Qwen3.5 397B A17B, had a 89% hallucination rate | Mar 2026: Still #5 of 8 with 1,449; the top four labs sat within 22 Arena points3Alibaba · 2 Aug 2026: Qwen 3.8 Max scored 158.0 on the Epoch Capabilities Index, its best modelMistral AI · 13 Jun 2023: Closed a funding round at a $260M post-money valuation, raising $113M in equityMistral AI · Oct 2023: Placed #5 of 5 on the Text Arena with 1,129 points, its first month in this seriesMistral AI · 11 Dec 2023: Closed a funding round at a $2.0B post-money valuation, raising $415M in equity | 11 Dec 2023: Mixtral 8x7B released, trained with an estimated 7.7×10²³ FLOP | Dec 2023: Rose to #3 of 5 on the Text Arena, 104 points behind OpenAI3Mistral AI · Jan 2024: Rose to #2 of 6 on the Text Arena, 104 points behind OpenAIMistral AI · 27 Feb 2024: Closed a funding round at a $2.0B post-money valuation, raising $16M in equityMistral AI · Mar 2024: Fell to #3 of 7 on the Text Arena, 78 points behind AnthropicMistral AI · Apr 2024: Fell to #4 of 7 on the Text Arena, 91 points behind OpenAIMistral AI · May 2024: Had about 35 staff | May 2024: Fell to #5 of 7 on the Text Arena, 99 points behind OpenAIMistral AI · 11 Jun 2024: Closed a funding round at a $6.0B post-money valuation, raising $500M in equity | Jun 2024: Fell to #7 of 7 on the Text Arena, 96 points behind OpenAIMistral AI · Jul 2024: Rose to #5 of 7 on the Text Arena, 36 points behind OpenAIMistral AI · Aug 2024: Fell to #6 of 8 on the Text Arena, 30 points behind OpenAIMistral AI · Oct 2024: Fell to #7 of 8 on the Text Arena, 75 points behind OpenAIMistral AI · Dec 2024: Fell to #8 of 8 on the Text Arena, 97 points behind OpenAIMistral AI · May 2025: Rose to #6 of 8 on the Text Arena, 111 points behind GoogleMistral AI · 2025 study: Mistral Medium 3.1 fell from 90.0% on standard Slovenian to 53.2% on the Cerkno dialect | Jun 2025: Fell to #7 of 8 on the Text Arena, 88 points behind GoogleMistral AI · 9 Sep 2025: Closed a funding round at a $13.7B post-money valuation, raising $2.0B in equity | Sep 2025: Had about 350 staffMistral AI · 31 Jan 2026: Annualized revenue reached $400M, from $100M on 11 Jun 2025Mistral AI · 2026 test: Its most reliable model on AA-Omniscience, Mistral Large 3, had a 84% hallucination rate | Mar 2026: Still #7 of 8 with 1,416; the top four labs sat within 22 Arena pointsMistral AI · 15 Sep 2026: Closed a funding round at a $24.0B post-money valuation, raising $3.5B in equity | 8 Sep 2026: Annualized revenue reached $1B, from $400M on 31 Jan 2026Mistral AI · 6 Oct 2026: Mistral Large 4 scored 150.6 on the Epoch Capabilities Index, its best modelNvidia · 14 Jun 2024: Nemotron-4 340B released, trained with an estimated 1.8×10²⁵ FLOPNvidia · 12 Oct 2024: Llama-3.1-Nemotron-70B-Instruct released, trained with an estimated 7.9×10²⁴ FLOPNvidia · 2025: Released 3 notable AI models in 2025, #10 by countNvidia · 2026 test: Its most reliable model on AA-Omniscience, NVIDIA Nemotron Nano, had a 83% hallucination rateNvidia · 4 Jun 2026: Nemotron 3 Ultra scored 146.0 on the Epoch Capabilities Index, its best model | 4 Jun 2026: Nemotron 3 Ultra released, trained with an estimated 6.6×10²⁴ FLOPNvidia · 2017–2026: Nvidia's A100 has trained 88 notable AI models since 2017, the most of any chip2023202420252026
Open AI Companies →
Coming up in AIDates already set

Dates already fixed in law, in an order or by the company: rules that start to apply, models that are switched off, agreements that begin. Each opens its official source or the Record thread that tracks it.

All dates ahead →

How every number is checked

Editorial standards →
01 · Sourced

One named source per chart

Each chart cites the report and the original survey, index or leaderboard behind it, with a link.

02 · Checked

Every number matched to the data

Figures in the text are checked against the chart's CSV before it is published, and re-checked on updates.

03 · Open

Yours to download and cite

Every chart comes with its CSV, a PNG and a ready-made citation, so you can verify it yourself.

Copied with its source. Want the next number like this in your inbox?Get the evidence