Skip to content

Market Research

AI Model Training Cost: Frontier Training Spend Data

Published November 13, 2026

OpenAI and Anthropic now spend billions of dollars a year to train AI models: OpenAI's research and training compute reached a reported $8.3 billion in 2025, and Anthropic's research and development expense reached $19.2 billion in the first nine months of 2026. The final training run for a single frontier model costs a few hundred million dollars, and most of the rest pays for experiments, research runs, and models that are never released.

Our analysts reviewed the 223 model-level training cost estimates in Epoch AI's database of AI models (downloaded November 13, 2026), Anthropic's S-1 filings, reported OpenAI and Anthropic compute budgets, and every training cost figure that a lab has published itself, to understand what frontier models cost to train as of the fourth quarter of 2026. We recorded each figure with its model or company, date, pricing method, and source, which gave a ledger of 258 dated entries.

Key Findings

  • Grok 4, released by xAI in July 2025, has the highest estimated final training run cost on record at $388 million in 2023 dollars, followed by OpenAI's GPT-4.5 at $366 million.
  • OpenAI's internal plans project $32 billion of training compute in 2026 and about $65 billion in 2027, up from $8.3 billion of R&D compute in 2025.
  • Anthropic spent 4.0 times as much on research and development as on serving customers in the first nine months of 2026, up from 2.3 times in 2025.
  • The cost of the most expensive training run released each year rose about 90 times from 2019 to 2025, an average of 2.1 times a year.
  • The same Grok 3 training run is estimated at $218 million on an amortized basis, $850 million at cloud rental prices, and $4.45 billion as the purchase price of its hardware.

A training run is the weeks or months in which a lab's computers process training data to produce a model such as GPT-4.5, Gemini, or Claude. Labs rarely say what a run cost. Most published figures are estimates built from the number and type of chips used, the length of the run, and a price per chip-hour, and the result depends heavily on which price is used. A company's yearly training budget is also many times larger than the cost of the runs behind its released models, because labs spend most of their compute on experiments first.

In the following sections, we compare the most expensive training runs on record, the estimates for the models behind ChatGPT, Gemini, and Claude, how the cost of the largest run has grown, three ways of pricing the same run, the cost per unit of compute, OpenAI and Anthropic's yearly training spending, what Anthropic's S-1 filing shows, and the training costs labs have published themselves.

The Most Expensive AI Training Runs on Record

We started with the single training runs that cost the most, since these are the figures most often quoted when people ask what it costs to build a frontier model. Epoch AI estimates each run's cost from its chip-hours, a depreciated price for the hardware, networking overhead, and the electricity used, and reports the result in 2023 dollars. Of the 223 models with an estimate, 34 cost more than $1 million to train, 15 cost more than $10 million, and 3 cost more than $100 million.

Horizontal bar chart of the 15 most expensive AI training runs by estimated amortized cost in 2023 dollars: Grok 4 $388 million, GPT-4.5 $366 million, Grok 3 $218 million, Llama 3.1-405B $52.9 million, Llama 4 Behemoth preview $44.6 million, GPT-4 $37.3 million, Grok-2 $31.6 million, Gemini 1.0 Ultra $30.7 million, Claude 3.5 Sonnet $25.9 million, Nemotron-4 340B $21.3 million, Mistral Large $14.1 million, Inflection-2 $13.5 million, A.X K2 $12.4 million, Inflection-2.5 $11.8 million, Falcon-180B $10.7 million
Highest Estimated Final Training Run Costs
Highest Estimated Final Training Run Costs
RankModelDeveloperReleasedTraining compute (FLOP)HardwareEstimated cost (2023 USD)
1Grok 4xAIJuly 20255.0 × 10²⁶200,000 chips$387.8 million
2GPT-4.5OpenAIFeb. 20253.8 × 10²⁶40,000 to 100,000 Nvidia H100 (estimated)$366.0 million
3Grok 3xAIFeb. 20253.5 × 10²⁶80,000 Nvidia H100$217.8 million
4Llama 3.1-405BMetaJuly 20243.8 × 10²⁵16,384 Nvidia H100$52.9 million
5Llama 4 Behemoth (preview)MetaApr. 20255.2 × 10²⁵32,000 Nvidia H100$44.6 million
6GPT-4OpenAIMar. 20232.1 × 10²⁵25,000 Nvidia A100$37.3 million
7Grok-2xAIAug. 20243.0 × 10²⁵Nvidia H100$31.6 million
8Gemini 1.0 UltraGoogle DeepMindDec. 20235.0 × 10²⁵57,000 Google TPU v4$30.7 million
9Claude 3.5 SonnetAnthropicJune 20242.7 × 10²⁵Not disclosed$25.9 million
10Nemotron-4 340BNvidiaJune 20241.8 × 10²⁵6,144 Nvidia H100$21.3 million
11Mistral LargeMistral AIFeb. 2024Not estimatedNvidia H100$14.1 million
12Inflection-2Inflection AINov. 20231.0 × 10²⁵5,000 Nvidia H100$13.5 million
13A.X K2SK TelecomJuly 20261.8 × 10²⁴Not disclosed$12.4 million
14Inflection-2.5Inflection AIMar. 20248.0 × 10²⁴Nvidia H100$11.8 million
15Falcon-180BTIISept. 20233.8 × 10²⁴4,096 Nvidia A100$10.7 million
16A.X K1SK TelecomDec. 2025Not estimated1,536 Nvidia H200$9.6 million
17Amazon TitanAmazonSept. 20234.8 × 10²⁴13,760 Nvidia A100$7.9 million
18DeepSeek-R1DeepSeekJan. 20253.3 × 10²⁴Not disclosed$6.8 million
19DeepSeek-V3DeepSeekDec. 20243.3 × 10²⁴2,048 Nvidia H800 SXM5$5.4 million
20SenseChatSenseTimeApr. 20233.9 × 10²⁴Not disclosed$5.3 million
  • US developers built 165 of the 223 models with a cost estimate, or 74%, and 13 of the 20 most expensive runs. Grok 4 cost about 10.4 times as much to train as GPT-4, which was released 28 months earlier.
  • Three runs from 2025, Grok 4, GPT-4.5, and Grok 3, are the only ones above $100 million. Each used around 3.5 × 10²⁶ to 5 × 10²⁶ FLOP of training compute, more than 10 times GPT-4's 2.1 × 10²⁵ FLOP.
  • OpenAI's GPT-6 Astra, released in September 2026, is the largest training run in Epoch's database at an estimated 9.4 × 10²⁶ FLOP on at least 100,000 Nvidia GB200 chips. Epoch has not yet published a cost estimate for it.

Training Cost Estimates for ChatGPT, Gemini, and Claude Models

Readers most often want the figure for the models behind ChatGPT, Gemini, and Claude, so we pulled every OpenAI, Google, and Anthropic model in Epoch's database with a cost or compute estimate. The newest models from all three companies have no estimate, because the labs stopped disclosing chip counts, run lengths, and parameter counts, which the Stanford AI Index 2026 also notes.

Scatter chart on a log scale of estimated training costs for OpenAI, Google, and Anthropic models from 2016 to 2025: OpenAI Five $4.3 million, GPT-3 $2.1 million, GPT-3.5 $4.9 million, PaLM $3.1 million, PaLM 2 $5.0 million, Claude 2 $4.9 million, GPT-4 $37.3 million, Gemini 1.0 Ultra $30.7 million, Claude 3.5 Sonnet $25.9 million, GPT-4.5 $366 million
Estimated Training Cost and Compute by Model
Estimated Training Cost and Compute by Model
ModelDeveloperReleasedTraining compute (FLOP)Estimated cost (2023 USD)
GPT-3 175BOpenAIMay 20203.1 × 10²³$2.1 million
GPT-3.5OpenAIMar. 20222.6 × 10²⁴$4.9 million
GPT-4OpenAIMar. 20232.1 × 10²⁵$37.3 million
GPT-4.5OpenAIFeb. 20253.8 × 10²⁶$366.0 million
GPT-5OpenAIAug. 20256.6 × 10²⁵No estimate published
GPT-6 AstraOpenAISept. 20269.4 × 10²⁶No estimate published
PaLM (540B)GoogleApr. 20222.5 × 10²⁴$3.1 million
PaLM 2GoogleMay 20237.3 × 10²⁴$5.0 million
Gemini 1.0 UltraGoogle DeepMindDec. 20235.0 × 10²⁵$30.7 million
Claude 2AnthropicJuly 20233.9 × 10²⁴$4.9 million
Claude 3.5 SonnetAnthropicJune 20242.7 × 10²⁵$25.9 million
Claude 3.7 SonnetAnthropicFeb. 20253.4 × 10²⁵No estimate published
  • GPT-4.5 cost about 9.8 times as much to train as GPT-4 on Epoch's estimate, and used about 18 times the compute. OpenAI's GPT-5 used an estimated 6.6 × 10²⁵ FLOP, less than GPT-4.5, because OpenAI put more of its effort into post-training.
  • Epoch's $25.9 million estimate for Claude 3.5 Sonnet is consistent with Anthropic CEO Dario Amodei's January 2025 statement that the model cost "a few $10M's" to train.
  • Google's Gemini 1.0 Ultra, released in December 2023, is the newest Gemini model with a cost estimate. Epoch has compute estimates for none of the Gemini 2.5, 3, or 4 models.

How Fast the Cost of the Largest Training Run Has Grown

Looking at only the most expensive run each year shows how quickly the top end has moved, without the many small research models that fill out the database. We took the highest estimate among models released in each year from 2016 to 2025.

Bar chart on a log scale of the most expensive training run released each year: GNMT $201,000 in 2016, AlphaGo Master $471,000 in 2017, ResNeXt-101 $134,000 in 2018, OpenAI Five $4.3 million in 2019, GPT-3 $2.1 million in 2020, Megatron-Turing NLG $3.8 million in 2021, GPT-3.5 $4.9 million in 2022, GPT-4 $37.3 million in 2023, Llama 3.1-405B $52.9 million in 2024, Grok 4 $388 million in 2025
Most Expensive Training Run Released Each Year
Most Expensive Training Run Released Each Year
YearModels with a cost estimateMost expensive modelDeveloperEstimated cost (2023 USD)
20167GNMTGoogle$201,332
20176AlphaGo MasterDeepMind$471,445
201812ResNeXt-101 32x48dFacebook$134,077
201922OpenAI FiveOpenAI$4.3 million
202027GPT-3 175BOpenAI$2.1 million
202142Megatron-Turing NLG 530BMicrosoft, Nvidia$3.8 million
202232GPT-3.5OpenAI$4.9 million
202345GPT-4OpenAI$37.3 million
20249Llama 3.1-405BMeta$52.9 million
20257Grok 4xAI$387.8 million
  • Epoch's own trend line across 45 frontier models puts the growth in final-run costs at 2.4 times a year since 2016. At that rate, Epoch projected in 2024 that the largest runs would cost more than $1 billion by 2027.
  • The number of models with a cost estimate fell from 45 in 2023 to 9 in 2024 and 7 in 2025, because fewer labs disclose the chip counts and run lengths an estimate needs.
  • The step from 2024 to 2025 was the largest in the series, a 7.3 times increase, as xAI trained Grok 3 on 80,000 Nvidia H100 chips and Grok 4 on a cluster of about 200,000 chips.

Three Ways to Price the Same Training Run

Published training costs for the same model can differ by a factor of 20, and the gap comes almost entirely from the pricing method. Epoch prices each run three ways: the share of the hardware's value used up during the run plus electricity, what the same chip-hours would cost to rent from a cloud provider, and the full purchase price of the hardware. We compared all three for every model where Epoch published them.

Grouped bar chart on a log scale of training cost by pricing method: GPT-3 $2.1 million amortized, $4.5 million cloud, $232 million upfront; GPT-4 $37.3 million, $81.4 million, $806 million; Gemini 1.0 Ultra $30.7 million, $206 million, $620 million; Llama 3.1-405B $52.9 million, $176 million, $928 million; Grok 3 $218 million, $850 million, $4.45 billion
Final Training Run Cost by Pricing Method
Final Training Run Cost by Pricing Method
ModelAmortized hardware and energy (Epoch)Cloud rental (Epoch)Upfront hardware (Epoch)Cloud rental (Stanford AI Index 2024)
GPT-3 175B$2.1 million$4.5 million$232.2 millionNot estimated
GPT-4$37.3 million$81.4 million$806.2 million$78 million
Gemini 1.0 Ultra$30.7 million$206.1 million$619.5 million$191 million
Llama 3.1-405B$52.9 million$176.0 million$928.4 millionNot estimated
Grok 3$217.8 million$849.7 million$4.45 billionNot estimated
Llama 4 Behemoth (preview)$44.6 millionNot estimated$1.78 billionNot estimated
Grok 4$387.8 millionNot estimated$10.72 billionNot estimated
  • For GPT-4, the cloud rental estimate is 2.2 times the amortized figure and the hardware purchase price is 21.6 times. The $78 million and $191 million figures most often quoted for GPT-4 and Gemini 1.0 Ultra are cloud rental estimates from the 2024 AI Index.
  • The hardware behind Grok 4 would cost an estimated $10.72 billion to buy, which reflects xAI's cluster of about 200,000 chips. Labs reuse that hardware for many runs, which is why the amortized estimate assigns only a fraction of the price to any one run.
  • Epoch's 2024 breakdown for GPT-4 and Gemini Ultra found that hardware made up 47% to 67% of total model development cost, research staff 29% to 49% including equity, and energy 2% to 6%.

Training Cost per Unit of Compute

Dividing each run's cost by its compute shows whether training is getting cheaper per calculation, which separates the effect of better chips from the effect of bigger runs. We calculated the cost per 10²⁴ floating-point operations (FLOP) for every model in the database that used at least that much compute.

Scatter chart of estimated training cost per 10 to the 24th FLOP from 2021 to 2025: GPT-3.5 $1.90 million, GPT-4 $1.78 million, Falcon-180B $2.86 million, Gemini 1.0 Ultra $0.61 million, Claude 3.5 Sonnet $0.96 million, Llama 3.1 405B $1.39 million, DeepSeek-R1 $2.05 million, GPT-4.5 $0.96 million, Grok 3 $0.62 million, Grok 4 $0.78 million
Estimated Cost per 10²⁴ FLOP for Runs Above 10²⁵ FLOP
Estimated Cost per 10²⁴ FLOP for Runs Above 10²⁵ FLOP
ModelDeveloperReleasedHardwareCost per 10²⁴ FLOP (2023 USD)
GPT-4OpenAIMar. 2023Nvidia A100$1.78 million
Inflection-2Inflection AINov. 2023Nvidia H100$1.34 million
Gemini 1.0 UltraGoogle DeepMindDec. 2023Google TPU v4$0.61 million
Nemotron-4 340BNvidiaJune 2024Nvidia H100$1.18 million
Claude 3.5 SonnetAnthropicJune 2024Not disclosed$0.96 million
Llama 3.1-405BMetaJuly 2024Nvidia H100$1.39 million
Grok-2xAIAug. 2024Nvidia H100$1.07 million
Grok 3xAIFeb. 2025Nvidia H100$0.62 million
GPT-4.5OpenAIFeb. 2025Nvidia H100$0.96 million
Llama 4 Behemoth (preview)MetaApr. 2025Nvidia H100$0.86 million
Grok 4xAIJuly 2025Not disclosed$0.78 million
  • GPT-4.5 cost an estimated $0.96 million per 10²⁴ FLOP, 46% less than GPT-4's $1.78 million. GPT-4.5 was trained on Nvidia H100 chips, and GPT-4 on the older A100.
  • Google's runs on its own TPU chips have some of the lowest estimates, $0.61 million for Gemini 1.0 Ultra, though Epoch prices TPUs from an estimated production cost rather than a market price, which makes those figures less certain.
  • The cost per unit of compute fell much more slowly than total cost rose. Grok 4's run cost 0.44 times as much per FLOP as GPT-4 but about 10 times as much in total, because it used about 24 times the compute.

OpenAI and Anthropic Spending on Training and Research Compute

We then moved from single runs to company budgets, because a lab's yearly spending on training and research compute is the larger and more consequential number. OpenAI's figures come from its internal financial documents and from court testimony by its president, Greg Brockman. Anthropic's come from internal documents and from its S-1 filing with the SEC. The subscriber revenue behind ChatGPT is covered in our ChatGPT paid conversion rate statistics.

Bar chart of yearly spending in billions of dollars: OpenAI R&D compute $4 billion in 2024 and $8.3 billion in 2025, projected training compute $32 billion in 2026 and $65 billion in 2027; Anthropic R&D expense $0.42 billion in 2023, $2.27 billion in 2024, $6.34 billion in 2025, and $19.24 billion in the first nine months of 2026
Reported Compute Spending at OpenAI and Anthropic
Reported Compute Spending at OpenAI and Anthropic
CompanyYearMeasureAmountBasis
OpenAI2022Computing and data$416 millionReported
OpenAI2024Training compute$3.0 billionReported
OpenAI2024Research compute amortization$1.0 billionReported
OpenAI2024Inference compute$1.8 billionReported
OpenAI2025R&D compute (training and research)$8.3 billionReported
OpenAI2025Inference compute$8.0 billionReported
OpenAI2026All computing power$50 billionProjection, Brockman testimony
OpenAI2026Training compute$32 billionInternal projection
OpenAI2027Training computeAbout $65 billionInternal projection
Anthropic2023Cloud compute, all uses$285 millionReported
Anthropic2024Training compute$1.5 billionReported
Anthropic2024Compute and infrastructure, all uses$2.52 billionS-1 filing
Anthropic2025Training and research compute$4.1 billionReported
Anthropic2025Compute and infrastructure, all uses$7.33 billionS-1 filing
Anthropic2026, first nine monthsCompute and infrastructure, all uses$18.81 billionS-1 filing
  • OpenAI's R&D compute grew 2.1 times from 2024 to 2025, and its 2026 training projection is 3.9 times the 2025 figure. Brockman testified that OpenAI's computing costs rose from about $30 million in 2017 to a projected $50 billion in 2026.
  • Epoch estimated that the final training runs of OpenAI's released models, including GPT-4.5 at about $400 million, accounted for about 10% of the company's roughly $5 billion of R&D compute in 2024. The rest went to research, experimental runs, and unreleased models.
  • Anthropic's S-1 does not split training from other research compute, because it runs its compute as one shared pool across all workloads and buys most of it 12 to 18 months in advance under multi-year contracts.

What Anthropic's S-1 Filing Shows About Training Spend

Anthropic is the only frontier lab whose training spending appears in audited financial statements, through the S-1 registration it filed ahead of its planned IPO and amended on November 2, 2026. We took its research and development expense, which the filing says consists primarily of compute for model development, research, and internal use of Claude, and set it against the cost of serving paying customers. The power behind that compute is covered in our AI electricity demand forecasts.

Grouped bar chart of Anthropic research and development expense against cost of revenue, in millions of dollars: 2023, $422 and $26; 2024, $2,275 and $752; 2025, $6,338 and $2,773; first nine months of 2025, $3,594 and $1,796; first nine months of 2026, $19,240 and $4,812
Anthropic Research Spending, Serving Costs, and Compute
Anthropic Research Spending, Serving Costs, and Compute
PeriodRevenueResearch and developmentCost of revenueCompute and infrastructure, all usesR&D to cost of revenue
2023$41 million$422 million$26 million$433 million16.2 times
2024$386 million$2,275 million$752 million$2,522 million3.0 times
2025$4,585 million$6,338 million$2,773 million$7,326 million2.3 times
First nine months of 2025$2,545 million$3,594 million$1,796 million$4,381 million2.0 times
First nine months of 2026$33,242 million$19,240 million$4,812 million$18,808 million4.0 times
Third quarter of 2026$16,956 million$9,499 million$2,432 million$9,863 million3.9 times
  • Anthropic's research and development expense rose 5.4 times year over year in the first nine months of 2026, while revenue grew 13 times, so research spending fell from 141% of revenue to 58%.
  • Compute and infrastructure for all workloads was 57% of revenue in the first nine months of 2026, down from 160% in 2025. Anthropic reports compute commitments of about $895.4 billion, of which about $69.8 billion falls due in the 12 months after September 30, 2026.
  • The filing warns that a frontier training run can take weeks or months and cost many millions of dollars, and that Anthropic has had failed or underperforming runs in the past.

Training Costs Labs Have Published Themselves

Labs that publish a dollar figure almost always price their own chip-hours at a rental rate, and none of these figures has been checked by an outside party. We collected every one we could find, along with the statements OpenAI and Anthropic executives have made about their own models and the most-cited outside estimates for OpenAI's largest runs. US labs come first, followed by labs in China for comparison.

Dot and range chart on a log scale of published training costs: GPT-6 Astra $540 million to $1.01 billion (Wccftech), GPT-4.5 $386 million (Epoch AI), GPT-4 more than $100 million (Sam Altman), Claude 3.5 Sonnet $25.9 million (Epoch AI), DeepSeek-V3 $5.6 million, BLOOM-176B $2 million to $5 million, OLMo 3 Think 32B $2.75 million, Xiaomi MiMo-V2.6 Pro RL run $2.62 million, MiniMax-M1 RL run $534,700, DeepSeek-R1 reasoning stage $294,000
Published Training Cost Figures for Individual Models
Published Training Cost Figures for Individual Models
ModelDeveloperPublishedFigureWho stated itWhat it covers
GPT-6 AstraOpenAISept. 2026$540 million to $1.008 billionWccftech estimate100,000 GPUs for 90 to 120 days at $2.50 to $3.50 per GPU-hour
GPT-4.5OpenAIOct. 2025$386 million (90% interval $170 million to $890 million)Epoch AI estimateFinal run at cloud rental prices
GPT-4OpenAIApr. 2023More than $100 millionSam Altman, OpenAI CEOTraining, no breakdown given
Claude 3.5 SonnetAnthropicJan. 2025"A few $10M's"Dario Amodei, Anthropic CEOTraining, no exact figure given
OLMo 3 Think 32BAi2Dec. 2025$2.75 millionLab paper1,024 H100s for 56 days at $2 per hour, full pipeline
BLOOM-176BBigScienceNov. 2022$2 million to $5 millionModel cardCloud equivalent, including preliminary experiments
DeepSeek-V3DeepSeekDec. 2024$5.576 millionLab paper2.788 million H800 hours at $2 per hour, final run only
DeepSeek-R1DeepSeekSept. 2025$294,000Lab paper in NatureReasoning training stages on top of V3
MiniMax-M1MiniMaxJune 2025$534,700Lab paperReinforcement learning run only
MiMo-V2.6 Flash and ProXiaomiSept. 2026About $850,000 and $2.62 millionRelease noteReinforcement learning post-training only
  • No lab has published a dollar cost for any current ChatGPT, Gemini, or Claude model, and the current model cards from OpenAI, Anthropic, and Google DeepMind give no compute, chip-hour, or cost figure.
  • The lab-published figures leave out most development spending. Meta has said the full cost of its OPT-175B model was roughly twice its final run, and the BLOOM team found its final run produced about 37% of the project's emissions.
  • OpenAI has said GPT-6 Astra was trained on more than 100,000 GPUs at its Abilene, Texas, data center, and Nvidia CEO Jensen Huang has said a coming OpenAI run will use 400,000.

What Frontier Training Spend Data Shows

The cost of a single frontier run has moved from tens of millions of dollars in 2023 to several hundred million in 2025, but the company budgets have grown much faster. OpenAI's projected $32 billion of training compute for 2026 would pay for about 87 runs the size of GPT-4.5 at Epoch's amortized estimate, and most of it is planned for experiments and research rather than the runs behind released models.

The figures are also becoming harder to check. Epoch could estimate costs for 45 models released in 2023 and 7 released in 2025, and none of the three companies behind ChatGPT, Gemini, and Claude publishes training compute for its current models. Anthropic's S-1 is now the most detailed public record, and it reports research spending as one pool rather than run by run.

In the third quarter of 2026 alone, Anthropic's research and development expense was $9.5 billion, about 26 times Epoch's amortized estimate for the entire GPT-4.5 training run.

Data: ai-model-training-cost-data-2026.csv

Data Collection and Methodology

Source 1 is our own research: a ledger of 258 dated training cost entries compiled on November 13, 2026. It contains 223 model-level estimates from Epoch AI's AI Models database, downloaded on November 13, 2026 with entries through November 10, 2026, plus 35 company spending, filing, and published run cost figures. Each row records the model or measure, organization, date, training compute where available, value, currency, pricing basis, and source type. The ledger is available as a CSV download.

Model-level costs are Epoch AI's estimates of the amortized hardware and energy cost of each model's final training run, expressed in 2023 US dollars. Epoch multiplies training chip-hours by a depreciated hardware price, adds 23% for cluster networking, and adds energy cost. Where Epoch also publishes cloud rental and upfront hardware estimates, we show them as published. We removed two DeepSeek re-releases that repeat their parent model's compute and cost. Costs per 10²⁴ FLOP divide each cost estimate by Epoch's training compute estimate.

OpenAI has not published its compute spending. Its figures come from internal financial documents reported by The Information and The New York Times, compiled in Epoch AI's AI Companies data, and from Greg Brockman's May 2026 court testimony reported by Reuters. Anthropic figures marked S-1 come from its Form S-1/A filed November 2, 2026; third-quarter 2026 values subtract the first-half figures in its original S-1 from the nine-month totals. Anthropic's research and development expense includes personnel and other costs as well as compute.

Lab-published figures are listed as stated, with the pricing assumption each lab used. We calculated products, differences, and ratios from the published values. Each percentage was rounded once, from unrounded inputs. All money figures are in US dollars.

Sources

  1. Top GEO Agencies, AI model training cost ledger (Source 1, our research), compiled Nov. 13, 2026. Download the data
  2. Epoch AI, AI Models database (all models and frontier models files), downloaded Nov. 13, 2026, CC BY 4.0. epoch.ai
  3. Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, and David Owen, How Much Does It Cost to Train Frontier AI Models?, Epoch AI, 2024. epoch.ai
  4. Josh You, Most of OpenAI's 2024 compute went to experiments, Epoch AI, Oct. 10, 2025. epoch.ai
  5. Epoch AI, Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won't), 2025. epoch.ai
  6. Epoch AI, AI Companies data, compute spend file, downloaded Nov. 13, 2026. epoch.ai
  7. Anthropic, PBC, Form S-1/A, filed Nov. 2, 2026. sec.gov
  8. Anthropic's S-1/A puts its compute commitments at $895.4 billion, sysnote, Nov. 2026. sysnote.net
  9. Anthropic's IPO filing shows most of its compute goes to R&D, Tom's Hardware, 2026. tomshardware.com
  10. OpenAI projects $50 billion spending on computing power this year, Brockman says, Reuters, May 5, 2026. reuters.com
  11. OpenAI adds $111 billion to its cash burn forecast as AI costs spiral beyond projections, The Decoder, Feb. 2026. the-decoder.com
  12. OpenAI Boosts Revenue Forecasts, Predicts $111 Billion More Cash Burn Through 2030, The Information, Feb. 20, 2026. theinformation.com
  13. Stanford HAI, Artificial Intelligence Index Report 2026, Apr. 2026. hai.stanford.edu
  14. Stanford HAI, Artificial Intelligence Index Report 2024, Chapter 1: Research and Development. hai.stanford.edu
  15. Will Knight, OpenAI's CEO Says the Age of Giant AI Models Is Already Over, Wired, Apr. 17, 2023. wired.com
  16. Dario Amodei, On DeepSeek and Export Controls, Jan. 2025. darioamodei.com
  17. Digital Applied, census of training cost figures published by AI labs, Sept. 22, 2026. digitalapplied.com
  18. OpenAI's ~$1 Billion GPT-6 Astra Heralds The Age Of AGI, According To NVIDIA's Jensen Huang, Wccftech, Sept. 2026. wccftech.com
  19. Top GEO Agencies, ChatGPT Paid Conversion Rate: Subscriber Statistics Over Time. topgeoagencies.net
  20. Top GEO Agencies, AI Electricity Demand Forecasts: Every Published Estimate. topgeoagencies.net

Source: Top GEO Agencies analysis of Epoch AI, AI Models database, downloaded Nov. 13, 2026.

Source: Top GEO Agencies analysis of Epoch AI, Frontier AI Models data; Stanford HAI, AI Index Report 2024.

Source: Top GEO Agencies analysis of Epoch AI, AI Companies compute spend data (reports by The Information, The New York Times, and Fortune); Reuters, May 5, 2026; The Decoder, Feb. 2026; Anthropic S-1/A, Nov. 2, 2026.

Source: Anthropic, PBC, Form S-1/A, filed Nov. 2, 2026; third-quarter values calculated from the nine-month and first-half figures.

Source: Top GEO Agencies analysis of lab papers, model cards, and release notes compiled by Digital Applied (Sept. 2026); Wired, Apr. 2023; Dario Amodei, Jan. 2025; Epoch AI, Oct. 2025; Wccftech, Sept. 2026.