State of AI Tokens 2026
What a million tokens costs, how many tokens the platforms process, and what each token returns, year by year from 2022 to 2026 to date
Thank you for reading this post, don't forget to subscribe!
- Google said it processed more than 3.2 quadrillion tokens a month across its products in May 2026, about 330 times the 9.7 trillion it processed in spring 2024.
- At GPT-3.5 quality, $1 bought about 50,000 tokens in November 2022 and about 14.3 million in October 2024, a 286-fold gain on Stanford HAI’s price series (Report AI calculation); at the higher GPT-4 bar, $1 bought about 5 million tokens in October 2026.
- At GPT-4 (March 2023) quality, $1 bought about 5 million tokens in October 2026 on a 3:1 input:output blend, about 188 times the 26,667 that GPT-4 itself bought at launch (Report AI calculation from list prices).
- At OpenAI’s flagship list price, $1 bought about 250,000 tokens in October 2026, against about 229,000 for GPT-4o in 2024 (Report AI calculation).
- Silicon Data’s usage-weighted LLM Token Expenditure Index fell below $1 per million tokens by 22 October 2026, about 54% under its late-May 2026 peak of $2.07, according to Axios.
- OpenRouter said it routed 790 trillion tokens in October 2026, 25 times its January volume, and that AI agents consumed 76% of them.
- Combined OpenAI and Anthropic run-rate revenue rose about 18-fold between December 2024 and the third quarter of 2026, while Google’s monthly token volume rose about 330-fold (Report AI comparison).
- Wharton’s October 2026 survey found 80% of US enterprise leaders reporting a positive return on AI but only 28% a strongly positive one; no company discloses its return per token.
Download the data (CSV) · 26 metrics with source, date and confidence for every value. Free to quote with attribution and a link.
- 1. Price per token — A dollar buys about 188 times more GPT-4-level output than at GPT-4’s March 2023 launch, but at OpenAI’s frontier it buys only about 9% more than in 2024
- 2. The market price — The price buyers actually paid nearly doubled to $2.07 per million tokens by late May 2026, then fell below $1 as work moved to cheaper models
- 3. Volume — Disclosed token volumes rose by two orders of magnitude in two years, and ByteDance now reports more tokens a month than Google
- 4. The Jevons test — Token volume rose about 330-fold and the two leading labs’ revenue about 18-fold while the price of GPT-4-level output fell by more than half: cheaper tokens enlarged the bill
- 5. Return per token — Providers earn roughly $0.70 to $1.00 per million tokens routed, and fewer than a third of buyers report strong returns
- Comparison
- The numbers, 2024–2026
- Outlook: what to watch in 2027
- FAQ · About the research
A dollar buys about 188 times more GPT-4-level output than at GPT-4’s March 2023 launch, but at OpenAI’s frontier it buys only about 9% more than in 2024
There is no single price of a token, and most confusion about AI costs comes from comparing different prices. We track three, back to 2022 where prices exist. The first holds capability fixed: what is the cheapest way to buy output as good as GPT-4 was when it launched in March 2023? The second follows the frontier: what does the newest flagship model cost at list price? The third, covered in chapter 2, is the average price the market actually pays, weighted by usage. All three are expressed per million tokens and converted to tokens per dollar using a blend of three input tokens to one output token, which is close to the mix of a typical chat or coding request. The blend matters because output tokens cost four to eight times more than input tokens on most price lists.
The longest series starts at GPT-3.5 quality: Stanford HAI’s price fell from $20 per million tokens in November 2022, about 50,000 tokens per dollar, to $0.07 in October 2024, about 14.3 million, a 286-fold gain in two years. At the fixed GPT-4 bar the gain is also large and mostly in the past. GPT-4 listed at $30 per million input tokens and $60 per million output tokens in March 2023, a blended $37.50, or about 26,667 tokens per dollar. By December 2024 DeepSeek-V3, which reports a higher MMLU score than GPT-4, listed at $0.27 and $1.10: about 2.1 million tokens per dollar, 78 times more. DeepSeek’s September 2025 price raised that to about 3.2 million, and by October 2026 small models from OpenAI and Anthropic at $0.10 and $0.50 reached 5 million. That is 188 times the March 2023 launch figure (75 times GPT-4 Turbo’s November 2023 figure), but only 2.4 times the 2024 figure. Our series uses named models with a published MMLU above GPT-4’s, so it is a conservative floor; smaller models that pass the bar on other tests are cheaper still.
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|
| GPT-3.5-level (Stanford series) | 50,000 (Nov) | 800,000 (gpt-3.5-turbo, Nov) | 14.3M (Oct) | – | – |
| GPT-4-level, cheapest named model | no such model | 66,667 (GPT-4 Turbo); 26,667 at Mar launch | 2.09M (DeepSeek-V3) | 3.17M (DeepSeek-V3.2) | 5.0M (GPT-6 Luna / Haiku 5.5) |
| Blended price at GPT-4 level, $ per 1M | – | $15.00 ($37.50 at launch) | $0.48 | $0.32 | $0.20 |
| OpenAI flagship | 50,000 (Davinci) | 66,667 (GPT-4 Turbo) | 228,571 (GPT-4o) | 290,909 (GPT-5) | 250,000 (GPT-6.1 Sol) |
| Anthropic flagship | no public API | 83,333 (Claude 2) | 33,333 (Claude 3 Opus) | 100,000 (Opus 4.5) | 125,000 (Opus 5.5) |
Notes: All values DERIVED: blended price = (3 x input price + output price) / 4; tokens per $1 = 1,000,000 / blended price. GPT-3.5 level = Stanford HAI’s MMLU 64.8 series ($20.00 Nov 2022, $0.07 Oct 2024); 2023 is OpenAI’s gpt-3.5-turbo-1106 list price ($1 / $2), and Stanford’s series has no 2025 or 2026 update. GPT-4 level = published MMLU at or above GPT-4’s 86.4. 2022 OpenAI price ($20 per 1M, input = output) from its 1 Sep 2022 Davinci price cut (MEDIUM). Anthropic’s flagship was Claude 2 in 2023 and the Opus tier from 2024. Anthropic 2026 prices verified on Anthropic’s pricing page on 1 Oct 2026 (HIGH); 2026 OpenAI prices MEDIUM. Excludes batch, cache and negotiated discounts.REPORTAI report-ai.org
The frontier tells a different story. OpenAI’s top model bought about 50,000 tokens per dollar in 2022 (GPT-3 Davinci at $20 per million) and, after GPT-4’s premium launch, about 66,667 by the end of 2023 (GPT-4 Turbo). It then bought about 229,000 tokens per dollar in 2024 (GPT-4o at $2.50 and $10), about 291,000 in 2025 (GPT-5 at $1.25 and $10) and about 250,000 in October 2026 (GPT-6.1 Sol at $2 and $10). Within 2026 the range was wide: GPT-5.6 Sol listed at $5 and $30 in July, only 88,889 tokens per dollar, before OpenAI halved prices in September. Anthropic’s Opus tier moved from 33,333 tokens per dollar for Claude 3 Opus in 2024 to 100,000 for Opus 4.5 in late 2025 and 125,000 for Opus 5.5 in 2026, while its top Fable 5.1 tier sits at 50,000. Google went the other way, scheduling Gemini 4 Argon’s price to double to $4 and $20 on 30 November 2026.
The pattern is consistent with how a16z and Epoch AI describe the market. A given capability falls in price quickly once several vendors and open-weight models can supply it: a16z put the rate at about ten times a year in 2024, Epoch measured a median of about 50 times a year in early 2025, and Epoch’s 2026 work implies about 13 times a year. New frontier capability is priced for what buyers will pay and gets cheaper only when competitors catch up. So a claim that AI is getting cheaper is true at a fixed capability level and largely untrue at the frontier. These are list prices; negotiated enterprise discounts, batch rates and cached-input pricing, which Anthropic lists at $0.10 to $0.25 per million, can lower actual bills considerably.
View the data
| OpenAI flagship | Anthropic flagship | |
|---|---|---|
| 2022 | 50 | |
| 2023 | 66.7 | 83.3 |
| 2024 | 229 | 33.3 |
| 2025 | 291 | 100 |
| 2026 | 250 | 125 |
Notes: DERIVED from list prices (see data file metrics openai_flagship_tokens_per_dollar and anthropic_flagship_tokens_per_dollar). Models: Davinci / text-davinci-003 (2022), GPT-4 Turbo (2023), GPT-4o (2024), GPT-5 (2025), GPT-6.1 Sol (Oct 2026); Claude 2 (2023), Claude 3 Opus (2024), Opus 4.5 (2025), Opus 5.5 (2026). Anthropic had no public API in 2022. The GPT-3.5-level series (50,000 in 2022, 800,000 in 2023, 14.3 million in 2024) spans 286x and is shown in the table above rather than on this axis. 2026 is to date; within 2026 OpenAI’s flagship ranged from 88,900 (GPT-5.6 Sol, Jul) to 250,000.REPORTAI report-ai.org
View the data
| Event | Group | Start | End |
|---|---|---|---|
| Davinci cut to $20/M | OpenAI | 2022-09 | |
| ChatGPT launch (free) | OpenAI | 2022-11 | |
| gpt-3.5-turbo $2/M | OpenAI | 2023-03 | |
| GPT-4 at $30/$60 | OpenAI | 2023-03 | |
| Claude 2 at $8/$24 | Anthropic | 2023-07 | |
| GPT-4 Turbo $10/$30 | OpenAI | 2023-11 | |
| GPT-4o at $2.50/$10 | OpenAI | 2024-08 | |
| DeepSeek-V3 launch | DeepSeek | 2024-12 | |
| GPT-5 at $1.25/$10 | OpenAI | 2025-08 | |
| DeepSeek V3.2 cut | DeepSeek | 2025-09 | |
| Opus 4.5 at $5/$25 | Anthropic | 2025-11 | |
| SDLLMTK index launch | Market | 2025-12 | |
| GPT-5.6 Sol $5/$30 | OpenAI | 2026-07 | |
| GPT-6 Sol/Luna -50% | OpenAI | 2026-09 | |
| Opus 5.5 at $4/$20 | Anthropic | 2026-09 | |
| Haiku 5.5 $0.10/$0.50 | Anthropic | 2026-10 | |
| Argon intro $2/$10 | 2026-10 | 2026-11 |
Notes: Gemini 4 Argon’s introductory price of $2 input / $10 output runs from general availability on 13 Oct 2026 to 30 Nov 2026, then $4 / $20 (MEDIUM). GPT-5.6 Sol price from Report AI page 964 (MEDIUM). DeepSeek V3.2-Exp: $0.28 / $0.42, Sep 2025. SDLLMTK launch month from secondary reporting. 2022-2023 events: OpenAI’s Davinci price cut (1 Sep 2022), ChatGPT release (30 Nov 2022), gpt-3.5-turbo (1 Mar 2023), Claude 2 (11 Jul 2023), GPT-4 Turbo at DevDay (6 Nov 2023); GPT-4’s own launch (14 Mar 2023) is also shown.REPORTAI report-ai.org
A token is a chunk of text, roughly three-quarters of an English word. Providers charge separately for input tokens (what you send) and output tokens (what the model writes), and output is usually four to eight times dearer. To turn two prices into one we use a blend of three input tokens to one output token: blended price = (3 x input + output) / 4. Tokens per dollar is one million divided by that blended price. A different blend changes the level but rarely the direction. Reasoning models add a third kind of token, hidden ‘thinking’ output billed at the output rate, which is why a cheap per-token price can still mean an expensive task. Tokenizers also differ: the same paragraph can be 30% more tokens on one model than another.
The price buyers actually paid nearly doubled to $2.07 per million tokens by late May 2026, then fell below $1 as work moved to cheaper models
List prices say what a token could cost. The usage-weighted price says what buyers paid on average, given the models they chose. The only public series is Silicon Data’s LLM Token Expenditure Index, ticker SDLLMTK, launched in December 2025 and published daily as a blended dollar rate per million tokens across more than 200 models. Its launch level is reported at about $1.09. It rose to a peak of about $2.07 around 28 May 2026, fell about 20% by early July, stood at $1.16 to $1.18 in early August, reached a record low of $0.97 on 31 August or 1 September, and was below $1 again, at about $0.95 by our calculation, when Axios reported on 22 October. Exact values are hard to pin down: sources give the peak as $2.07, $2.12 or $2.05, and the timing of the break below $1 is reported inconsistently.
The fall was mostly a change in what buyers bought. A usage-weighted index declines when list prices are cut, but also when customers move work from expensive models to cheaper ones, and Silicon Data’s own research lead attributed much of the decline to routing away from frontier models toward open-weight ones. Its April 2026 segment readings show why mix matters so much: $4.20 per million tokens for frontier models against $0.85 for open-weight models, a fivefold gap. List-price cuts added to the fall: OpenAI cut prices on two GPT-5.6 models in late July and halved GPT-6 Sol and Luna in September, and Anthropic’s Haiku 5.5 launched in October at a tenth of Haiku 4.5’s price. The rise to May is less well explained in the sources; heavier use of frontier models for coding agents is the usual explanation.
View the data
| SDLLMTK, usage-weighted | |
|---|---|
| Dec 2025 | 1.09 |
| 28 May 2026 | 2.07 |
| Early Jul 2026 | 1.62 |
| 7 Aug 2026 | 1.17 |
| 31 Aug 2026 | 0.97 |
| 21 Sep 2026 | 1 |
| 22 Oct 2026 | 0.95 |
Notes: Irregular dated readings only; no points are interpolated. Dec 2025 (about $1.09) is implied by a third-party explainer (MEDIUM, single source). Early July value ($1.62) reported as about 20% below the May peak. 7 Aug is the midpoint of $1.16-1.18 reported for 6-8 Aug. 21 Sep: ‘about $1’, 52% below peak (Tokenpost). 22 Oct (about $0.95) is DERIVED: 2.07 x (1 – 0.54). Peak CONTESTED ($2.07 vs $2.12 vs $2.05). Silicon Data has not been read directly. The index mixes price changes with model switching.REPORTAI report-ai.org
Cheaper tokens did not make tasks cheaper, because tasks used more tokens. In the a16z and OpenRouter study of 100 trillion tokens, the average prompt grew from about 1,500 tokens to more than 6,000 between November 2024 and November 2025, and the average completion from about 150 to 400. At OpenAI’s flagship list prices, that average request cost about $0.0053 in November 2024 and $0.0115 a year later, more than double, while the blended token price fell 21%. Reasoning models widen the gap. In April 2025 Artificial Analysis spent about $2,767 running OpenAI’s o1 through a seven-benchmark suite against $109 for GPT-4o, and in October 2026 its index cost $10.04 per task on Claude Opus 5.5 at maximum effort against $1.27 on GPT-6.1 Sol, a 7.9-fold gap between models whose list prices differ by about two times.
For a buyer, the unit that matters is the cost of a completed task, and that depends on three things that move independently: the price per token, the number of tokens the model spends, and how often the work has to be redone. Only the first is published. OpenAI’s GDPval study of September 2025 is the closest thing to a task-level comparison with human labour: it found frontier models completed expert tasks about 100 times faster and 100 times cheaper than the experts, but it counted only inference time and API cost, not review, iteration or integration. We have not found a study that prices the same task at 2024 and 2026 token rates alongside a stated wage, so the return on a task cannot yet be tracked year by year.
| Indicator | 2024 | 2025 | 2026 |
|---|---|---|---|
| Average request at OpenAI flagship list price | $0.0053 (Nov) | $0.0115 (Nov) | not measured |
| Average prompt / completion on OpenRouter, tokens | ~1,500 / ~150 | ~6,000 / ~400 | not published |
| Benchmark cost, reasoning vs non-reasoning model | – | $2,767 (o1) vs $109 (GPT-4o), suite | $10.04 vs $1.27 per task (Opus 5.5 vs GPT-6.1 Sol) |
| AI vs expert cost on GDPval tasks | – | about 100x cheaper (inference only) | – |
Notes: Average request cost is DERIVED: prompt tokens x input price + completion tokens x output price, using GPT-4o ($2.50 / $10) for 2024 and GPT-5 ($1.25 / $10) for 2025. The 2025 and 2026 benchmark figures are on different bases (whole suite vs per task) and are not comparable with each other. GDPval excludes human review and rework.REPORTAI report-ai.org
The fixed-capability price answers ‘what does a given level of quality cost now?’ It falls fastest. The flagship list price answers ‘what does the best model cost?’ It moves in steps and has risen at times. The usage-weighted price answers ‘what did buyers pay on average?’ It moves with both prices and choices, so it can fall even if no price list changes. A widely shared headline in 2026 read a falling usage-weighted index as falling demand. It is not a demand measure: OpenRouter’s volume rose 37% in October while the index sat near its low. Before quoting any token price, check which of the three it is.
Disclosed token volumes rose by two orders of magnitude in two years, and ByteDance now reports more tokens a month than Google
Google publishes the longest cross-product series. It processed about 9.7 trillion tokens a month in spring 2024, 480 trillion in May 2025, 980 trillion in July 2025 and 1.3 quadrillion in October 2025, and said at I/O in May 2026 that it had passed 3.2 quadrillion, about seven times a year earlier. Since then it has reported only its first-party API traffic: about 7 billion tokens a minute in the third quarter of 2025, 16 billion in the first quarter of 2026, 22 billion in the second and nearly 32 billion in the third, more than four times a year earlier. At that rate the API alone handles about 1.4 quadrillion tokens a month. All of these figures are self-reported and unaudited, and Google counts tokens across products as different as Search’s AI Overviews and the Gemini app.
The fastest-growing large series is Chinese. ByteDance’s Volcano Engine said its Doubao models processed more than 4 trillion tokens a day in December 2024, about 50 trillion in December 2025, 120 trillion in March 2026 and more than 180 trillion in June 2026. Converted to a 30-day month, June’s figure is 5.4 quadrillion tokens, more than Google’s last all-product figure, although Doubao’s count includes ByteDance’s internal use and its own consumer app. China’s National Data Administration put national use at about 100 billion tokens a day in early 2024, about 100 trillion by the end of 2025 and more than 140 trillion in March 2026, with 21.1 quadrillion tokens for 2025 as a whole. The 2024 baseline is disputed: another official account puts the start of 2025 at more than 1 trillion a day.
View the data
| Google, all products | |
|---|---|
| Apr-May 2024 | 9.7 |
| May 2025 | 480 |
| Jul 2025 | 980 |
| Oct 2025 | 1,300 |
| May 2026 | 3,200 |
Notes: Plotted only at dated disclosures; uneven spacing. The range is about 330x, so early points sit near zero on a linear axis. Self-reported, unaudited (MEDIUM). Google has not published an all-product monthly figure since May 2026; its Q3 2026 API rate (about 32B tokens per minute) is a different measure and is not plotted.REPORTAI report-ai.org
OpenRouter, which routes developers’ requests to more than 300 models, shows where the growth comes from. Its weekly volume rose from about 10 billion tokens in January 2024, on a single-source baseline, to more than 7 trillion in December 2025 and about 194 trillion in the week of 26 October 2026, according to a chart posted by its investor Menlo Ventures. OpenRouter’s own October recap reports 790 trillion tokens for the month, 25 times January, of which agents consumed 76%, against about 8% a year earlier. An agent calls a model repeatedly, feeding back its own output and tool results, so tokens per task rise with every step it takes without a person. OpenRouter’s traffic is weighted toward developers and open-weight models and is not representative of the whole market.
The largest Western API businesses are missing from these series. OpenAI said its API processed more than 6 billion tokens a minute in October 2025, up from about 300 million in 2023, but gave no figure at DevDay 2026. Microsoft processed more than 100 trillion tokens in the first quarter of 2025 and over 500 trillion through Foundry in fiscal 2025, then switched to counting customers on track to use a trillion tokens each. Anthropic does not publish token volume. So no market total can be built. Our floor adds only the two firms with comparable all-product figures, Google and ByteDance: about 130 trillion tokens a month in 2024, 2.8 quadrillion in late 2025 and 8.6 quadrillion in mid-2026, a 66-fold rise that excludes most of the market.
| Platform (unit) | 2024 | 2025 | 2026 to date |
|---|---|---|---|
| Google, all products (per month) | 9.7T (Apr-May) | 1.3 quadrillion (Oct) | 3.2 quadrillion (May) |
| Google, model APIs (per minute) | – | 10B+ (Q4) | ~32B (Q3) |
| OpenAI API (per minute) | – | 6B+ (Oct) | not disclosed |
| Microsoft (per quarter) | ~20T (Q1, derived) | 100T+ (Q1) | not disclosed |
| ByteDance Doubao (per day) | 4T+ (Dec) | ~50T (Dec) | 180T+ (Jun) |
| China, national (per day) | ~0.1T (early, contested) | ~100T (end) | 140T+ (Mar) |
| OpenRouter (per week) | ~10B (Jan) | ~7T (Dec, derived) | ~194T (late Oct) |
| Google + ByteDance floor (per month) | ~130T | ~2.8 quadrillion | ~8.6 quadrillion |
Notes: Units differ by row and must not be added across rows. Microsoft 2024 = 100T / 5 (DERIVED from ‘up 5x’). OpenRouter 2025 = more than 1T per day x 7 (DERIVED); 2026 weekly figure is from an investor’s chart. The floor row is DERIVED: Google monthly + Doubao daily x 30; it excludes OpenAI, Anthropic, Microsoft, Alibaba and others, and excludes OpenRouter to avoid double counting. China’s 2024 baseline is CONTESTED. All volumes MEDIUM unless marked.REPORTAI report-ai.org
A token is not a fixed unit. Tokenizers split text differently, so the same document can produce noticeably different counts on different models. Platforms also count different things: Google sums tokens across products from Search to Gemini; ByteDance includes its own apps and internal use; OpenRouter counts paid traffic only and excludes free model variants; Microsoft has switched from token totals to counts of large customers. Images, audio and video are converted to tokens at rates each company sets. None of these figures is audited. Compare each platform with itself over time, not with another platform, and treat any cross-platform total as a floor.
Token volume rose about 330-fold and the two leading labs’ revenue about 18-fold while the price of GPT-4-level output fell by more than half: cheaper tokens enlarged the bill
In 1865 the economist William Stanley Jevons observed that more efficient steam engines increased Britain’s coal consumption rather than reducing it, because cheaper power found new uses. The same question applies to tokens: when the price of a unit falls, does total revenue fall, or does demand grow faster than the price drops? We built three indexes with 2024 set to 100. Price uses the cheapest GPT-4-level blended list price from chapter 1. Volume uses Google’s monthly tokens in May of each year, the only cross-product series measured in the same month each year. Revenue uses the combined annualised run-rate revenue of OpenAI and Anthropic, the two labs that hold most enterprise API spend, measured the same way in each year. All three are DERIVED, and each is a proxy for a market nobody measures in full. The question matters because it decides whether falling prices shrink or grow the token business.
The result is a strong Jevons pattern. The price index fell from 100 to 66 in 2025 and 42 in 2026. Volume rose to about 4,950 in 2025 and 32,990 in 2026, roughly 330 times its 2024 level. Revenue rose to 446 in 2025 and 1,769 in 2026: combined OpenAI and Anthropic run-rate revenue went from about $6.5 billion in December 2024 to about $29 billion at the end of 2025 and about $115 billion by the third quarter of 2026, about 18 times. Every fall in price was matched many times over by growth in tokens consumed. Revenue therefore grew faster than the GPT-4-level price fell, but far more slowly than volume. Menlo Ventures’ enterprise spending estimates are not used here: its September 2026 figure of an annualised $176.8 billion uses a new definition that is not comparable with its $11.5 billion for 2024 or its $37 billion, restated to $33 billion, for 2025.
View the data
| Price (GPT-4 level) | Revenue, OpenAI + Anthropic (run-rate) | Volume (Google) | |
|---|---|---|---|
| 2024 | 100 | 100 | 100 |
| 2025 | 66 | 446 | 4,948 |
| 2026 | 42 | 1,769 | 32,990 |
Notes: All DERIVED. Price = cheapest GPT-4-level blended list price ($0.4775, $0.315, $0.20 per 1M). Revenue = combined OpenAI + Anthropic annualised run-rate revenue: about $6.5B (Dec 2024: OpenAI ~$5.5B + Anthropic ~$1B), about $29B (end 2025: ~$20B + ~$9B), about $115B (OpenAI ~$50B end Sep 2026 + Anthropic $65B+ end Jul 2026); run rates are unaudited and include subscriptions. Volume = Google monthly tokens in May (9.7T, 480T, 3,200T). Implied revenue per token = revenue index / volume index x 100: 100, 9.0, 5.4. Menlo’s enterprise spend is not used: its 2026 figure ($176.8B annualised) is on a new definition, not comparable with $11.5B (2024) or $37B, restated $33B (2025). On a linear axis the price bars are barely visible beside volume; read the values. Base year: 2024 = 100 is kept because the volume leg has no 2023 or 2022 value on the same definition (Google’s first monthly disclosure is spring 2024; no other platform published comparable figures). The price leg has a 2023 value ($15.00 at GPT-4 Turbo, index 3,141 on the 2024 base) and the revenue leg has one (about $1.7B, index 26), but a base year needs all three.REPORTAI report-ai.org
Dividing the revenue index by the volume index gives an implied revenue per token, and it fell to about 9.0 in 2025 and 5.4 in 2026: roughly 95% less money per token than in 2024. That is much steeper than the fall in the GPT-4-level list price, and the difference is the mix. Most new tokens are consumed by cheap models, by long prompts that are largely cached input, and by free or bundled consumer products whose tokens earn nothing directly. The usage-weighted price in chapter 2 shows the same mix effect within 2026. For providers this is the central tension of the token business: revenue rises with volume, but each token earns less, so margins depend on serving costs falling at least as fast as prices.
The index has clear limits. The revenue leg covers two labs only, uses unaudited run rates dated two months apart in 2026, and includes consumer subscriptions as well as token sales. Google’s volume includes consumer products whose tokens no one pays for directly, so revenue and volume do not cover the same activity. Menlo’s spending series cannot fill the gap: its 2024 base excludes inference, its 2025 figure was restated from $37 billion to $33 billion, and its 2026 figure is an annualised run rate on a new definition. A volume index built from Doubao, which rose from 4 trillion to 180 trillion tokens a day between December 2024 and June 2026, would show a similar order of magnitude. The direction is robust; the exact ratios are not, and should be quoted as orders of magnitude rather than as precise figures.
The index shows that falling token prices have not reduced providers’ total revenue; demand has grown far faster than prices fell. It cannot show whether the extra tokens are worth what is paid for them, because revenue measures what buyers pay, not the value they get. It also cannot separate genuine new demand from tokens generated by design choices: longer prompts, verbose reasoning and agents that retry. Some of the growth may be waste that falling prices made affordable. If prices stopped falling, part of that volume might disappear. Treat the index as evidence about the market’s size and direction, not about returns.
Providers earn roughly $0.70 to $1.00 per million tokens routed, and fewer than a third of buyers report strong returns
Return per token has no standard definition, and nobody publishes it. We built the measures the data allows and labelled each DERIVED. On the provider side, revenue per million tokens divides what customers pay by the tokens they consume. OpenRouter is the one platform where both can be estimated: Sacra put its gross spend at about $100 million a year in May 2025, when a16z said it handled more than 100 trillion tokens a year, which gives about $1.00 per million tokens. In August 2026 Sacra estimated its revenue at $160 million a year on a take rate of about 5%, implying about $3.2 billion of spend, while Menlo put annual volume at 4.5 quadrillion tokens, which gives about $0.71. The 2026 figure depends on the take rate: 4% gives $0.89, 6% gives $0.59.
Other providers allow only rough estimates. ByteDance’s Volcano Engine is reported to have earned about RMB 1.5 billion from model services in 2025. Its daily token count rose from about 4 trillion to 50 trillion during the year, about 8.6 quadrillion tokens in total on a linear path between readings, which works out to about RMB 0.17, or roughly $0.024, per million tokens. That figure understates external revenue per token because Doubao’s count includes ByteDance’s own apps. The US labs disclose revenue but not tokens. Anthropic’s S-1 reported $16.3 billion of revenue in the first half of 2026, about 80% of it from consumption-based billing per one reading of the filing, but no token volume. Silicon Data’s index, at about $0.95 in October 2026, is in effect the market’s average revenue per million tokens across the models it tracks.
| Measure | 2024 | 2025 | 2026 to date |
|---|---|---|---|
| OpenRouter spend per 1M tokens routed | – | ~$1.00 (mid-year) | ~$0.71 (Aug) |
| Silicon Data usage-weighted price per 1M | – | ~$1.09 (Dec launch) | ~$0.95 (22 Oct) |
| Volcano Engine revenue per 1M tokens | – | ~$0.024 (RMB 0.17) | – |
| Implied revenue per token (OpenAI + Anthropic), index | 100 | 9.0 | 5.4 |
| OpenAI + Anthropic run-rate revenue | ~$6.5B (Dec) | ~$29B (end year) | ~$115B (Jul-Sep) |
Notes: DERIVED. OpenRouter: 2025 = $100M GMV / 100T tokens per year; 2026 = ($160M revenue / 0.05 take rate) / 4.5 quadrillion tokens per year; the take rate is an estimate. Volcano Engine: RMB 1.5B / about 8.6 quadrillion tokens (Report AI estimate from 4T, 16.4T and 50T daily readings, linear between them), at about RMB 7.2 per dollar; includes internal tokens, so it understates. Revenue-per-token index = OpenAI + Anthropic run-rate revenue index / Google volume index x 100, from the Jevons exhibit. Run rates include consumer subscriptions and are not audited.REPORTAI report-ai.org
On the buyer side, the evidence depends on how the question is asked. Wharton and GBK Collective’s annual survey of US enterprise leaders found about 65% reporting a positive return in 2024, 74% in 2025 and 80% in October 2026, but only 28% describe the return as strongly positive. McKinsey’s 2026 survey found 37% of organisations reporting any EBIT contribution from AI, essentially unchanged from 2025, with about 6% attributing more than 5% of EBIT to it. MIT’s NANDA study reported in August 2025 that 95% of the generative-AI pilots it studied showed no measurable profit-and-loss impact within about six months, and IBM found that 25% of AI initiatives had met their expected ROI. Snowflake’s surveys report average returns of $1.41 per dollar in 2025 and $1.49 in 2026.
These studies are CONTESTED rather than contradictory: they ask different people different questions over different periods. Their common finding is that a majority of firms report some return, a minority report a large one, and the share reporting a large return has not grown as fast as spending. None of them divides return by tokens consumed, and no company discloses firm-level return per token, so the buyer’s return per token cannot be measured with public data. The cost side is now visible enough to matter: about one in five organisations told McKinsey in 2026 that AI operating costs were constraining their use. That is the first sign that the token bill, rather than access to the models, has started to limit adoption.
| Survey (measure) | 2024 | 2025 | 2026 |
|---|---|---|---|
| Wharton / GBK (positive ROI) | ~65% | 74% | 80% |
| Wharton / GBK (strongly positive) | – | – | 28% |
| McKinsey (any EBIT contribution) | – | 39% | 37% |
| Writer (significant ROI) | – | – | 29% |
| IBM (initiatives meeting expected ROI) | – | 25% | – |
| MIT NANDA (pilots with measurable P&L) | – | 5% | – |
| Snowflake (US$ returned per $1) | – | $1.41 | $1.49 |
Notes: CONTESTED: questions, samples and success bars differ and the rows should not be compared with each other. Wharton 2024 (about 65%) is DERIVED from the 2025 report’s +9 pp change. McKinsey says its 2026 figure is essentially unchanged from 2025, which suggests the 39% for 2025 may have been RESTATED. MIT NANDA counted P&L impact within about six months. Snowflake 2026 is the cross-industry average (49% ROI) from 2,050 respondents.REPORTAI report-ai.org
Revenue per million tokens = customer spend through a platform / tokens consumed, in millions. For OpenRouter, spend = reported revenue / take rate (about 5%). Implied revenue per token index = (OpenAI + Anthropic run-rate revenue index / Google volume index) x 100, with 2024 = 100. Request cost = prompt tokens x input price + completion tokens x output price, at the flagship list price. Each measure uses public inputs, so each inherits their weaknesses: estimated take rates, unaudited token counts and spend figures with changing scope. We did not build a buyer-side return per token, because no published source gives both a firm’s measured return and its token consumption. That figure does not exist in public data.
From 2022 to 2026 the cost of fixed capability fell by two orders of magnitude, frontier tokens per dollar rose about fivefold, and volume and lab revenue are traceable only from 2023-24
Over the full period from 2022 to 2026 the long-run price story is the clearest. At GPT-3.5 quality, $1 bought about 50,000 tokens in November 2022, 800,000 in November 2023 and 14.3 million in October 2024, a 286-fold gain in two years, after which Stanford’s series has no update. At OpenAI’s flagship, tokens per dollar went from 50,000 in 2022 to 66,667, 228,571, 290,909 and 250,000: about a fivefold gain over four years, with a dip at GPT-4’s March 2023 launch, when a dollar bought only 26,667. Volume cannot be traced back as far. The first public token count is OpenAI’s 300 million a minute in 2023, and the cross-product series from Google and ByteDance start only in 2024. On revenue, OpenAI recognised about $28 million in 2022; the two leading labs’ combined run rate was about $1.7 billion at the end of 2023 and about $115 billion by the third quarter of 2026, but the 2022 figure is on a different basis and is not used for a multiple.
Taken year by year, the three lenses move at very different speeds. Tokens per dollar at the GPT-4 level rose from about 2.1 million in 2024 to 3.2 million in 2025 and 5 million in 2026, a gain of about 2.4 times; most of the improvement since 2023 happened before 2024. At OpenAI’s flagship, tokens per dollar were 229,000, 291,000 and 250,000, effectively flat. Volume moved fastest: Google’s monthly tokens grew about 49 times from May 2024 to May 2025 and about 7 times in the following year, and Doubao’s daily tokens grew about 12 times in 2025 and more than 3 times in the first half of 2026. Combined OpenAI and Anthropic run-rate revenue grew about 4.5 times in 2025 and about 4 times again by the third quarter of 2026; Menlo’s enterprise spend estimates changed definition in 2026 and give no comparable multiple.
Across platforms, growth rates differ by an order of magnitude, partly because of different starting points. OpenRouter grew fastest, from tens of billions of tokens a week to nearly 200 trillion, because it began small and caught the move to agents. Google’s growth slowed in relative terms as its base grew, from about 49 times in its first year to about 7 times in its second, while its API traffic still quadrupled year on year. Chinese volumes grew at least as fast as American ones on every series we found, and ByteDance’s single disclosed figure now exceeds Google’s. On revenue, the two leading US labs’ combined run rate grew about 18 times in under two years, against about 330 times for Google’s token volume, so revenue per token fell by about 95% even as total revenue soared.
| Indicator | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|
| Tokens per $1, GPT-3.5 level | 50,000 | 800,000 | 14.3M | – | – |
| Tokens per $1, GPT-4 level | no such model | 66,667 | 2.09M | 3.17M | 5.0M |
| Tokens per $1, OpenAI flagship | 50,000 | 66,667 | 228,571 | 290,909 | 250,000 |
| OpenAI API tokens per minute | – | ~300M | – | 6B+ | – |
| Google tokens per month | – | – | 9.7T | 1.3 quadrillion | 3.2 quadrillion |
| Doubao tokens per day | not launched | not launched | 4T+ | ~50T | 180T+ |
| OpenRouter tokens per week | not founded | – | ~10B | ~7T | ~194T |
| Enterprise GenAI spend (Menlo) | – | $1.7B (restated from $2.3B) | $11.5B | $37B (restated $33B) | $176.8B annualised, new definition (not comparable) |
| OpenAI + Anthropic revenue | ~$28M (OpenAI recognised; different basis) | ~$1.7B run rate | ~$6.5B | ~$29B | ~$115B |
| Leaders reporting positive ROI (Wharton) | – | – | ~65% | 74% | 80% |
Notes: Each cell is the latest reading in that calendar year; months differ by row (see data file). Tokens per dollar, run rates, OpenRouter 2025 and Wharton 2024 are DERIVED. Units differ by row. ‘-‘ means no sourced value. 2022 revenue is OpenAI’s full-year recognised revenue and is not comparable with the later run rates. Menlo’s 2023 figure is on its 2025-report basis (RESTATED from $2.3B); its 2026 figure uses a new definition and is not comparable with earlier totals (CONTESTED); no multiple is computed from that row. 2026 values are to date.REPORTAI report-ai.org
Year over year, every comparable measure
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | 2025→26 |
|---|---|---|---|---|---|---|
| Tokens per $1 at GPT-4 (March 2023) quality, cheapest named model | — | 66,667 | 2.09M | 3.17M | 5.0M | +58% |
| Tokens per $1 at OpenAI’s flagship list price | 50,000 | 66,667 | 228,571 | 290,909 | 250,000 | -14% |
| Tokens per $1 at Anthropic’s flagship list price (Opus tier from 2024) | — | 83,333 | 33,333 | 100,000 | 125,000 | +25% |
| Google: tokens processed per month across all products | — | — | 9.7T | 1.3 quadrillion | 3.2 quadrillion | +146% |
| OpenRouter: tokens routed per week | — | — | ~10B | ~7T | ~194T | +2,671% |
| ByteDance Doubao models: tokens processed per day | — | — | 4T+ | ~50T | 180T+ | +260% |
| China: average daily token calls, national | — | — | ~0.1T (100B) | ~100T | 140T+ | +40% |
| Disclosed tokens per month, Google plus ByteDance (cross-platform floor) | — | — | ~130T | ~2.8 quadrillion | ~8.6 quadrillion | +207% |
| Enterprise generative-AI spending (Menlo Ventures estimate; definition changed in 2026) | — | $1.7B | $11.5B | $37B | $176.8B (annualised) | +378% |
| OpenAI plus Anthropic revenue (combined annualised run rate; 2022 recognised revenue) | ~$28M (OpenAI recognised revenue, full year) | ~$1.7B | ~$6.5B | ~$29B | ~$115B | +297% |
| Jevons check: token volume index (Google monthly tokens, May each year), 2024 = 100 | — | — | 100 | 4,948 | 32,990 | +567% |
| Jevons check: revenue index, OpenAI + Anthropic run-rate revenue, 2024 = 100 | — | 26 | 100 | 446 | 1,769 | +297% |
| Enterprise leaders reporting positive ROI from generative AI (Wharton / GBK) | — | — | ~65% | 74% | 80% (28% strongly positive) | +8% |
- Tokens per dollar at GPT-4 quality rose from about 2.1 million (2024) to 3.2 million (2025) to 5.0 million (2026), against 27,000 for GPT-4 in March 2023, so most of the gain came before 2024. At OpenAI’s flagship the figure barely moved: 229,000, then 291,000, then 250,000.
- Google’s monthly token volume went from 9.7 trillion (spring 2024) to 480 trillion (May 2025) to 3.2 quadrillion (May 2026); Doubao’s daily volume went from 4 trillion (Dec 2024) to 50 trillion (Dec 2025) to 180 trillion (Jun 2026).
- Combined OpenAI and Anthropic run-rate revenue went from about $6.5 billion (Dec 2024) to $29 billion (end 2025) to $115 billion (Q3 2026): about 18 times in under two years, an order of magnitude less than token volume. Menlo’s enterprise spend estimates ($11.5 billion for 2024; $37 billion, restated to $33 billion, for 2025; an annualised $176.8 billion in Sep 2026 on a new definition) are not comparable across that break.
- Reported business return improved only slightly. Wharton’s positive-ROI share went from about 65% to 74% to 80%, McKinsey’s EBIT-contribution share held near 37-39%, and the share reporting strong or full returns stayed below one in three.
All eleven themes side by side: State of AI, year over year.
The numbers, 2022 to 2026
Every measure in this report with its value for each year, as-of date, source and confidence. Change compares 2025 with 2026 to date where both are numeric.
Price: tokens per dollar
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Tokens per $1 at GPT-3.5 quality (Stanford HAI series, MMLU 64.8) tokens per US dollar (3:1 blend where input and output differ) | 50,000 Nov 2022, Stanford HAI, AI Index 2025: GPT-3.5-level inference at $20.00 per 1M tokens (Nov 2022) DERIVED | 800,000 Nov 2023, OpenAI, gpt-3.5-turbo-1106 list price $1 input / $2 output per 1M (DevDay, 6 Nov 2023) DERIVED | 14.3M Oct 2024, Stanford HAI, AI Index 2025: $0.07 per 1M tokens (Oct 2024) DERIVED | — | — | |
| Tokens per $1 at GPT-4 (March 2023) quality, cheapest named model tokens per US dollar (3:1 input:output blend) | — | 66,667 Nov 2023, OpenAI, GPT-4 Turbo (gpt-4-1106-preview) list price $10 input / $30 output per 1M, DevDay 6 Nov 2023 DERIVED | 2.09M Dec 2024, DeepSeek API list price (DeepSeek-V3: $0.27 input / $1.10 output per 1M); Report AI calculation DERIVED | 3.17M Sep 2025, DeepSeek API list price (DeepSeek-V3.2-Exp: $0.28 input cache-miss / $0.42 output per 1M); Report AI calculation DERIVED | 5.0M Oct 2026 (to date), OpenAI GPT-6 Luna ($0.10 / $0.50, after the 22 Sep 2026 cut) and Anthropic Claude Haiku 5.5 ($0.10 / $0.50 for prompts up to 100K tokens, 7 Oct 2026), as reported; Report AI calculation DERIVED | +58% |
| Tokens per $1 at OpenAI’s flagship list price tokens per US dollar (3:1 blend) | 50,000 Sep-Dec 2022, OpenAI pricing: Davinci (GPT-3) cut from $0.06 to $0.02 per 1,000 tokens on 1 Sep 2022; text-davinci-003 (28 Nov 2022) at the same $0.02 DERIVED | 66,667 Nov 2023, OpenAI, GPT-4 Turbo list price $10 / $30 per 1M (DevDay, 6 Nov 2023) DERIVED | 228,571 Aug 2024, OpenAI API pricing, GPT-4o ($2.50 input / $10.00 output per 1M) DERIVED | 290,909 Aug 2025, OpenAI API pricing, GPT-5 ($1.25 input / $10.00 output per 1M) DERIVED | 250,000 Oct 2026 (to date), GPT-6.1 Sol at $2 / $10 (implied by GPT-6.1 Terra at $1 / $5 being half its price, 26 Oct 2026; GPT-6 Sol cut 50% to $2 / $10 on 22 Sep 2026), as reported by Technobezz and iClarified DERIVED | -14% |
| Tokens per $1 at Anthropic’s flagship list price (Opus tier from 2024) tokens per US dollar (3:1 blend) | — | 83,333 Jul-Nov 2023, Anthropic pricing, Claude 2 and Claude 2.1 ($8 input / $24 output per 1M) DERIVED | 33,333 Mar 2024, Anthropic pricing, Claude 3 Opus ($15 input / $75 output per 1M) DERIVED | 100,000 Nov 2025, Anthropic, Claude Opus 4.5 launch pricing ($5 / $25 per 1M) DERIVED | 125,000 Sep 2026 (to date), Anthropic pricing page, Claude Opus 5.5 ($4 / $20 per 1M), verified by Report AI on 1 Oct 2026 DERIVED | +25% |
Price: market
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Usage-weighted market price per million tokens (Silicon Data LLM Token Expenditure Index, SDLLMTK) US$ per 1M tokens | — | — | — | ~$1.09 Dec 2025 (launch), Implied by a third-party explainer saying the index nearly doubled from about 1.09 to about 2.12 between December 2025 and May 2026 MEDIUM | ~$0.95 22 Oct 2026 (to date), Axios, citing Silicon Data: index fell below $1, 54% under its late-May peak of $2.07 DERIVED | -13% |
| Annual decline in the price of a fixed level of capability times cheaper per year | — | — | ~10x per year Nov 2024, a16z, Welcome to LLMflation (GPT-3-level MMLU 42: $60 per 1M in Nov 2021 to $0.06 in Nov 2024, 1,000x in three years) MEDIUM | ~50x per year (median; range 9x-900x) Mar 2025, Epoch AI, LLM inference price trends MEDIUM | ~13x per year 2026 (to date), Epoch AI, The plunging price of thought: cost of a fixed performance level fell about 47% per quarter over three years DERIVED | -75% |
| Cost of running a standard benchmark on a frontier reasoning model US$ (see note: suite total in 2025, per task in 2026) | — | — | — | $2,767 (o1, seven-benchmark suite) vs $109 (GPT-4o) Apr 2025, Artificial Analysis via TechCrunch / DeepLearning.AI The Batch MEDIUM | $10.04 per task (Claude Opus 5.5, max) vs $1.27 (GPT-6.1 Sol, max) 28 Oct 2026 (to date), Artificial Analysis Intelligence Index v5, as reported by FourWeekMBA MEDIUM | -100% |
Volume: platforms
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Google: tokens processed per month across all products tokens per month | — | — | 9.7T Apr-May 2024, Alphabet (Sundar Pichai, Google I/O, May 2025, citing the year-earlier figure) MEDIUM | 1.3 quadrillion Oct 2025, Alphabet Q3 2025 earnings remarks MEDIUM | 3.2 quadrillion May 2026 (latest all-surface figure), Alphabet, Google I/O 2026 MEDIUM | +146% |
| Google: first-party model API tokens per minute tokens per minute | — | — | — | 10B+ Q4 2025, Alphabet Q4 2025 earnings remarks (series as tabulated by PPC Land) MEDIUM | ~32B Q3 2026 (28 Oct 2026, to date), Alphabet Q3 2026 earnings remarks: nearly 32 billion tokens per minute, 40% more than last quarter and more than 4x a year ago MEDIUM | +220% |
| OpenAI: API tokens per minute tokens per minute | — | ~300M 2023 (DevDay, Nov 2023 era), OpenAI, Sam Altman at DevDay 2025, recalling the 2023 figure MEDIUM | — | 6B+ 6 Oct 2025, OpenAI, DevDay 2025 (Sam Altman keynote) MEDIUM | — | |
| Microsoft: tokens processed per quarter (company-wide) tokens per quarter | — | — | ~20T Jan-Mar 2024, Microsoft FY25 Q3 call (‘over 100 trillion tokens this quarter, up 5x year over year’); Report AI calculation DERIVED | 100T+ Jan-Mar 2025, Microsoft FY25 Q3 earnings call MEDIUM | — | |
| OpenRouter: tokens routed per week tokens per week | — | — | ~10B Jan 2024, Crypto Briefing, citing OpenRouter’s growth chart (9,000x since January 2024) MEDIUM | ~7T Dec 2025, a16z / OpenRouter, State of AI: An Empirical 100 Trillion Token Study (more than 1 trillion tokens every day in the week before publication) DERIVED | ~194T week ending 26 Oct 2026 (to date), Chart posted by Menlo Ventures (an OpenRouter investor), as reported by FourWeekMBA and Crypto Briefing MEDIUM | +2,671% |
| ByteDance Doubao models: tokens processed per day tokens per day | — | — | 4T+ Dec 2024, Volcano Engine, Winter FORCE conference, as reported by SCMP MEDIUM | ~50T 18 Dec 2025, Volcano Engine president Tan Dai, FORCE conference MEDIUM | 180T+ Jun 2026 (latest found), Volcano Engine, FORCE conference (Doubao 2.1 launch), as reported by TMTPost and others MEDIUM | +260% |
| China: average daily token calls, national tokens per day | — | — | ~0.1T (100B) early 2024, National Data Administration (NDA), as reported by CnTechPost and China Daily CONTESTED | ~100T end 2025, NDA, National Data Resource Survey Report (2025) MEDIUM | 140T+ Mar 2026 (latest dated figure), NDA director Liu Liehong, policy briefing MEDIUM | +40% |
| Share of OpenRouter tokens consumed by AI agents % of tokens | — | — | — | ~8% Oct 2025, Secondary write-up of OpenRouter’s October 2026 recap (‘8% a year ago for agentic keys’) MEDIUM | 76% Oct 2026 (to date), OpenRouter, October 2026 data recap MEDIUM | +850% |
| Disclosed tokens per month, Google plus ByteDance (cross-platform floor) tokens per month | — | — | ~130T 2024 (May and Dec readings), Report AI calculation from Alphabet and Volcano Engine disclosures DERIVED | ~2.8 quadrillion Oct-Dec 2025, Report AI calculation DERIVED | ~8.6 quadrillion May-Jun 2026 (to date), Report AI calculation DERIVED | +207% |
Return: spend
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Enterprise generative-AI spending (Menlo Ventures estimate; definition changed in 2026) US$ per year | — | $1.7B 2023, Menlo Ventures, 2025 report footnote (2023 estimate excluding inference); originally $2.3B in its Nov 2024 report MEDIUM | $11.5B 2024, Menlo Ventures, 2025: The State of Generative AI in the Enterprise MEDIUM | $37B 2025, Menlo Ventures, Dec 2025 report MEDIUM | $176.8B (annualised) Sep 2026 (to date), Menlo Ventures, 2026: The State of Enterprise AI (3 Nov 2026) CONTESTED | +378% |
| Enterprise spending on LLM APIs (Menlo Ventures) US$ per year (annualised) | — | — | $3.5B Nov 2024, Menlo Ventures, 2025 Mid-Year LLM Market Update MEDIUM | $8.4B mid-2025, Menlo Ventures, 2025 Mid-Year LLM Market Update MEDIUM | — | |
| OpenAI plus Anthropic revenue (combined annualised run rate; 2022 recognised revenue) US$ per year (annualised) | ~$28M (OpenAI recognised revenue, full year) 2022, OpenAI recognised revenue for 2022 as reported in press coverage of its financials; Anthropic not disclosed MEDIUM | ~$1.7B Oct-Dec 2023, OpenAI annualised revenue passed $1.6B (The Information, via SiliconANGLE, 1 Jan 2024) + Anthropic about $100M (Epoch AI, dated 3 Oct 2023) DERIVED | ~$6.5B Dec 2024, OpenAI about $5.5B (CNBC, Jun 2025, citing the year-earlier figure) + Anthropic about $1B (press reports) DERIVED | ~$29B end 2025, OpenAI about $20B (page 935) + Anthropic about $9B (Bloomberg via TheNextWeb) DERIVED | ~$115B Jul-Sep 2026 (to date), OpenAI about $50B (end Sep 2026, CNBC) + Anthropic more than $65B (end Jul 2026, Bloomberg) DERIVED | +297% |
Return: Jevons
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Jevons check: token volume index (Google monthly tokens, May each year), 2024 = 100 index, 2024 = 100 | — | — | 100 Apr-May 2024, Report AI calculation from Alphabet disclosures DERIVED | 4,948 May 2025, Report AI calculation DERIVED | 32,990 May 2026, Report AI calculation DERIVED | +567% |
| Jevons check: revenue index, OpenAI + Anthropic run-rate revenue, 2024 = 100 index, 2024 = 100 | — | 26 Oct-Dec 2023, Report AI calculation DERIVED | 100 Dec 2024, Report AI calculation from press-reported run rates DERIVED | 446 end 2025, Report AI calculation DERIVED | 1,769 Jul-Sep 2026 (to date), Report AI calculation DERIVED | +297% |
Return: per token
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Revenue (spend) per million tokens routed, OpenRouter US$ per 1M tokens | — | — | — | ~$1.00 mid-2025, Sacra (OpenRouter GMV about $100M annualised, May 2025) and a16z (more than 100T tokens per year, mid-2025); Report AI calculation DERIVED | ~$0.71 Aug 2026 (to date), Sacra (OpenRouter revenue about $160M annualised, Aug 2026; take rate about 5%) and Menlo Ventures (4.5 quadrillion tokens per year run rate, 19 Aug 2026); Report AI calculation DERIVED | -29% |
Return: task economics
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Cost of an average API request at OpenAI’s flagship list price US$ per request | — | — | $0.0053 Nov 2024, a16z / OpenRouter 100 Trillion Token Study (average prompt about 1,500 tokens, completion about 150) at GPT-4o list price; Report AI calculation DERIVED | $0.0115 Nov 2025, Same study (average prompt about 6,000 tokens, completion about 400) at GPT-5 list price; Report AI calculation DERIVED | — |
Return: business evidence
| Measure | 2022 | 2023 | 2024 | 2025 | 2026 (to date) | Change |
|---|---|---|---|---|---|---|
| Enterprise leaders reporting positive ROI from generative AI (Wharton / GBK) % of respondents | — | — | ~65% 2024 survey, Wharton Human-AI Research / GBK Collective; Report AI calculation from the 2025 report’s ‘+9 pp’ change DERIVED | 74% Oct 2025, Wharton / GBK, Accountable Acceleration (2025 AI Adoption Report), 800+ US enterprise decision-makers MEDIUM | 80% (28% strongly positive) 13 Oct 2026 (to date), Wharton / GBK, 2026 AI Adoption Report, as reported by HPCwire, Unite.AI and Beri MEDIUM | +8% |
| Organisations reporting any EBIT contribution from AI (McKinsey State of AI) % of respondents | — | — | — | 39% 2025 survey (published Nov 2025), McKinsey, The State of AI 2025 MEDIUM | 37% survey 4 May – 8 Jun 2026, McKinsey, The state of AI in 2026: On the road to ROI (1,719 respondents, 97 countries) MEDIUM | -5% |
| Evidence of AI investments failing to return (share meeting expectations) % (definitions differ; see note) | — | — | — | 5% of pilots with measurable P&L impact (MIT NANDA); 25% of AI initiatives met expected ROI (IBM) Aug 2025 (MIT); May 2025 (IBM), MIT Project NANDA, The GenAI Divide (Aug 2025); IBM Institute for Business Value CEO Study (May 2025, 2,000 CEOs) CONTESTED | 29% report significant ROI (Writer); 5-10% report full payback (PYMNTS) 2026, Writer, Enterprise AI adoption in 2026; PYMNTS Intelligence (Jun 2026 survey of 60 executives) CONTESTED | +480% |
Numbers to stop quoting
| Widely repeated | What the evidence supports |
|---|---|
| “AI inference costs have fallen 280-fold (cited as the current price of frontier AI)” | The 280x figure is Stanford HAI’s measurement for GPT-3.5-level capability between November 2022 and October 2024. At OpenAI’s flagship list price, tokens per dollar went from about 229,000 in 2024 to about 250,000 in October 2026 (Report AI calculation). · detail |
| “The Silicon Data index shows demand for AI tokens is falling” | The LLM Token Expenditure Index is a usage-weighted price per million tokens, not a volume or spend measure. It fell 54% from late May to October 2026 while OpenRouter’s monthly volume rose 37% in October alone; Silicon Data itself has said the name misleads. · detail |
| “Google processes 1.3 quadrillion tokens a month” | True for October 2025; Google reported more than 3.2 quadrillion a month at I/O in May 2026. |
| “OpenRouter processes 10 quadrillion tokens a year” | That is one week’s volume (about 194 trillion, week of 26 October 2026) multiplied by 52, from a chart posted by an OpenRouter investor. OpenRouter’s own recap reports 790 trillion tokens for October 2026. |
| “95% of companies get no return from AI” | MIT NANDA’s 2025 figure is 95% of generative-AI pilots with no measurable P&L impact within about six months. Wharton’s 2026 survey has 80% of enterprise leaders reporting positive returns (28% strongly positive), and McKinsey has 37% reporting any EBIT contribution. · detail |
In 2027 the test is whether spending keeps pace with token volume as agents multiply the tokens behind every task
The data to date supports a simple account. Capability gets cheap once several suppliers can provide it, frontier capability stays expensive, and demand for tokens grows much faster than prices fall. If that continues, provider revenue keeps rising while revenue per token keeps falling, and the margin question moves to serving costs: chips, power and how efficiently models use tokens. Agents strengthen both effects. They raise the volume of tokens per task and shift work to cheaper models wherever a cheap model is good enough, which is what the usage-weighted price recorded in 2026.
Three things could change the picture. A sustained rise in frontier prices, of which Google’s scheduled Gemini 4 Argon increase is the first test, would show pricing power returning to the top of the market. A slowdown in volume growth below the rate of price decline would turn the Jevons pattern into falling total revenue. And evidence that buyers’ strong returns are not broadening, with the share stuck below a third while costs constrain one firm in five, would point to a correction in token demand driven by cost rather than capability. The indicators below are the ones we will track.
What to watch in 2027
| Indicator | Where it stands | What would change the picture |
|---|---|---|
| Silicon Data usage-weighted price (SDLLMTK) | about $0.95 per 1M tokens (22 Oct 2026); peak about $2.07 (late May 2026) | Sustained above $1.50 would mean buyers moving back to frontier models or frontier prices holding; below $0.50 would mean commoditisation is accelerating |
| Gemini 4 Argon price after its introductory period | $2 / $10 per 1M tokens, scheduled to rise to $4 / $20 on 30 Nov 2026 | If the rise holds and other labs follow, frontier tokens per dollar fall in 2027 for the first time since 2023 |
| Google monthly tokens across products | 3.2 quadrillion (May 2026), about 7x year on year | Growth below 3x at I/O 2027 would mark a clear slowdown; above 7x would mean agents are accelerating demand |
| Agent share of OpenRouter tokens | 76% (Oct 2026), from about 8% a year earlier | Above 85% would make agent loops the dominant driver of token demand; a fall would suggest agent workloads are being cut back on cost |
| Lab revenue vs token volume | Revenue index 1,769 (OpenAI + Anthropic run rate) vs volume index 32,990 (Google), 2024 = 100 | If 2027 run-rate revenue grows less than 2x while token volume grows more than 5x, revenue per token is falling faster than serving costs can easily follow |
| Share of leaders reporting strongly positive AI returns (Wharton) | 28% (Oct 2026) | Above 40% would show returns broadening; below 25% alongside rising cost constraints (McKinsey: about one in five firms) would point to a cost-driven pullback |
- Token
- The unit AI models read and write, roughly three-quarters of an English word on average. Models are priced and usage is measured in tokens, usually per million.
- Input and output tokens
- Input tokens are what is sent to the model (prompt, documents, conversation history); output tokens are what it generates. Output is typically priced four to eight times higher. Cached input is re-used input billed at a steep discount.
- Blended price (3:1)
- One price per million tokens combining input and output prices at three input tokens per output token: (3 x input + output) / 4. Report AI uses it for every tokens-per-dollar figure.
- Tokens per dollar
- 1,000,000 divided by the blended price per million tokens: how many tokens $1 buys at list price.
- Fixed-capability price
- The cheapest price at which a stated level of performance (here, GPT-4 as of March 2023) can be bought. It tracks how fast a given capability gets cheaper.
- Usage-weighted price
- The average price actually paid per million tokens, weighted by how much each model is used. Silicon Data’s LLM Token Expenditure Index (SDLLMTK) is the only public series; it falls when prices are cut or when buyers switch to cheaper models.
- Reasoning tokens
- Hidden intermediate output that reasoning models generate before answering, billed as output. They can multiply the tokens, and therefore the cost, of a task.
- Run rate
- Revenue annualised from a recent period, usually the latest month x 12. Not audited and not the same as calendar-year revenue.
- Take rate and GMV
- For a marketplace such as OpenRouter, GMV is the total customers spend through it and the take rate is the share it keeps. Revenue / take rate = GMV.
- Jevons paradox
- The observation, first made about coal in 1865, that making a resource cheaper or more efficient to use can increase total consumption of it. Applied here to tokens.
Frequently asked
How many tokens does $1 buy in 2026?
It depends on the model. At GPT-4 (March 2023) quality, OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 list at $0.10 per million input tokens and $0.50 per million output tokens as of October 2026, about 5 million tokens per dollar on a 3:1 input:output blend. OpenAI’s flagship GPT-6.1 Sol, at $2 and $10, buys about 250,000, and Anthropic’s Claude Opus 5.5, at $4 and $20, about 125,000 (Report AI calculations from list prices).
Are AI token prices still falling in 2026?
The average price paid is falling; frontier list prices mostly are not. Silicon Data’s usage-weighted LLM Token Expenditure Index fell from about $2.07 per million tokens in late May 2026 to below $1 by 22 October 2026, according to Axios, largely because buyers switched to cheaper models. OpenAI cut GPT-6 Sol by 50% in September 2026, while Google scheduled Gemini 4 Argon’s price to double to $4 and $20 on 30 November 2026.
How many tokens do AI platforms process?
Each platform uses its own unit. Google reported more than 3.2 quadrillion tokens a month across its products in May 2026, and its model APIs about 32 billion tokens a minute in Q3 2026. ByteDance reported more than 180 trillion tokens a day for its Doubao models in June 2026. OpenRouter routed 790 trillion tokens in October 2026. China’s National Data Administration put national use at more than 140 trillion tokens a day in March 2026. OpenAI, Microsoft and Anthropic have not published 2026 totals.
If tokens are cheaper, why are AI bills rising?
Each task now uses many more tokens. On OpenRouter the average prompt grew from about 1,500 to 6,000 tokens between November 2024 and November 2025 (a16z / OpenRouter), and reasoning models and agents generate far more output. At OpenAI’s flagship list prices, Report AI calculates that the average request cost about $0.0053 in November 2024 and $0.0115 in November 2025, more than double, even though the blended token price fell 21%.
What is the return per AI token?
No company or study publishes it. Report AI’s proxies: customers spent about $0.71 per million tokens routed through OpenRouter in August 2026, down from about $1.00 in mid-2025 (derived from Sacra and Menlo Ventures figures), and revenue per token is about 95% lower than in 2024 on our index (combined OpenAI and Anthropic run-rate revenue up about 18 times against Google’s token volume up about 330 times). On the buyer side, Wharton’s October 2026 survey found 80% of enterprise leaders reporting positive AI returns but only 28% strongly positive, and McKinsey’s 2026 survey found 37% reporting any EBIT contribution.
Figures are compiled from company disclosures, regulators, research firms and reputable reporting, and cross-checked against Report AI’s own index pages. Each value carries its as-of date, source and a confidence rating: HIGH where the primary source was read, MEDIUM where it rests on consistent reputable secondary reporting (most figures, because many primary pages cannot be opened from our research environment), DERIVED where it is our own arithmetic, and CONTESTED where credible sources disagree. A year means the calendar year the figure describes; 2026 values are the latest available and marked “to date”. Weekly and monthly user counts, and forecasts and outcomes, are never compared as if they were the same measure. Exhibits plot only dated, sourced readings; lines are not interpolated between them, and values with different definitions are never placed on one axis. Known gaps in this edition: No figure is rated HIGH apart from the Anthropic list prices verified for page 421. WebFetch could not open any page (blog.google and ppc.land failed DNS in this environment), so company releases, Silicon Data’s index page, Menlo’s PDF and OpenRouter’s recap were read only through search summaries.; Google has not published an all-products monthly token figure since May 2026; Q2 and Q3 2026 calls gave API tokens per minute only.; Microsoft has given no token total since fiscal 2025 (it now reports customers on track for 1 trillion tokens). OpenAI gave no API throughput at DevDay 2026; the ’15B+ per minute’ figure is unattributed. Anthropic does not disclose token volume (the S-1 coverage gives revenue mix only).; Not found: Alibaba Cloud Model Studio token volume, AWS Bedrock, Meta, xAI, Mistral, DeepSeek token totals.; Silicon Data’s SDLLMTK has no published month-end series in the sources found; 2026 readings are irregular, the launch value (about $1.09) is implied by a blog, no November 2026 value was found, and the peak ($2.07 vs $2.12 vs $2.05) and the timing of the drop below $1 are CONTESTED.; No usage-weighted market price exists for 2024 or most of 2025, so the Jevons price index uses the GPT-4-bar list price instead, which is a quality-adjusted list price, not the average price paid.; Epoch AI’s GPT-4-level price dataset could not be accessed; the GPT-4-bar series is built from named models’ list prices with a published MMLU above 86.4 and is a conservative floor.; Menlo’s 2026 report gives no dollar total for model-API spend, and its $176.8B annualised figure uses a new definition that is not comparable with its $11.5B (2024) and $37B / restated $33B (2025) totals, so Menlo cannot supply a 2024-to-2026 spend series. The Jevons revenue leg therefore uses OpenAI plus Anthropic run-rate revenue, which covers two labs only and includes subscriptions.; Revenue per million tokens can be estimated only for OpenRouter (dependent on an assumed ~5% take rate) and roughly for Volcano Engine (whose token count includes internal use). No lab discloses both revenue and tokens served.; No published study gives token counts for the same task at 2024 and 2026 prices alongside a human wage cost. GDPval (Sep 2025) is the only cost comparison with experts found, and it excludes oversight and rework. No firm-level return per token is disclosed anywhere.; 2022 and 2023: no token-volume disclosure was found from Google, Microsoft, ByteDance, Alibaba, Anthropic or OpenRouter; the only 2023 volume figure is OpenAI’s API at about 300 million tokens a minute, recalled in 2025. No usage-weighted price, revenue per token, request cost or ROI survey reading exists for either year. The Jevons comparison therefore stays at 2024 = 100: the price and revenue legs have 2023 values but the volume leg does not.; 2022 revenue is OpenAI’s recognised full-year revenue (about $28 million), not a run rate, and Anthropic’s 2022 revenue was not disclosed; Anthropic’s 2023 run rate (about $100 million) is dated October, not year end. Menlo’s 2023 LLM API spend figure was not found; Stanford’s GPT-3.5-level price series has no 2025 or 2026 update.. These reports compile and verify secondary data with original analysis; they are not original surveys. See our editorial standards, corrections log and fact-check register.
State of AI Usage · State of AI Spending · State of AI Finance · State of AI Work · State of AI Models & Prices · State of AI Tokens · State of AI Infrastructure & Energy · State of AI Geography · State of AI Industries · State of AI Risk · State of AI Regulation · Year over year · AI Market Pulse · AI statistics that are wrong