AI Price IndexThe economics of intelligence
Track how the cost of AI models, inference, compute and useful capability is changing, and what that implies for the cost of AI through 2030.
AI prices are increasingly difficult to describe with a single number. The cost of reproducing yesterday’s AI capability is collapsing, while new frontier models can remain expensive and worldwide spending on inference infrastructure continues to rise. This index tracks both sides of that equation.
Observed prices also reflect strategy. Providers are competing aggressively for developers, users and market share, so today’s API list price is not necessarily an equilibrium price. Three forces set the price customers pay:
- Technologypushes cost downward
- Demand and scarcitycan push price upward
- Competition and strategic subsidycan push customer prices below sustainable economics
The future price is the result of all three. Technology makes intelligence cheaper to produce. Competition can make it even cheaper to buy. But only the first of those effects is guaranteed to survive a normalization of market structure.
By Kim Salmi, SaaS Pricing Tracker. Last updated:
- Median frontier price
- $10/ 1M input tokens
- Median uncached input price across the tracked frontier basket.
- Cheapest frontier
- $2/ 1M input tokens
- Lowest-priced model currently classified as frontier in the maintained index.
- Cheapest budget
- $0.10/ 1M input tokens
- Lowest eligible budget input price in the maintained index.
- Fixed-capability decline
- ~13×cheaper a year
- Epoch AI’s measured historical decline in cost at a fixed level of benchmark performance since 2023.
Historical evidence, not a forward forecast
- Frontier premium
- 100×median frontier ÷ cheapest budget
- Price gap between the median tracked frontier tier and lowest tracked budget tier.
- 2030 central scenario
- ~100×cheaper
- AI Price Index central scenario for reproducing approximately 2026-level capability by 2030.
AI Price Index scenario
Our view: The price of individual tokens will fall much more slowly than the cost of obtaining a fixed level of intelligence. By 2030, today’s AI capabilities could cost around 1% of today’s amount even if the most capable models still command premium prices.
Exhibit 1
AI API prices collapsed after 2021, but the frontier is no longer following a simple downward curve
USD per million uncached input tokens; logarithmic scale
Note: Solid lines are observed list prices, held from one price change to the next. Fine dashed lines are a statistical trend that only continues the compound rate of the last 36 months. Shaded wedges are AI Price Index scenario ranges for 2028, widening to 2030. Forecasts start after the latest observation.
Show the data as a table
| Month | Median frontier | Cheapest frontier | Cheapest budget |
|---|---|---|---|
| Nov 2021 | $60 | $60 | $0.80 |
| Feb 2022 | $60 | $60 | $0.80 |
| May 2022 | $60 | $60 | $0.80 |
| Aug 2022 | $60 | $60 | $0.80 |
| Nov 2022 | $20 | $20 | $0.40 |
| Feb 2023 | $20 | $20 | $0.40 |
| May 2023 | $20 | $20 | $0.40 |
| Aug 2023 | $20 | $20 | $0.40 |
| Nov 2023 | $20 | $20 | $0.40 |
| Feb 2024 | $20 | $20 | $0.40 |
| May 2024 | $10 | $10 | $0.50 |
| Jun 2024 | $12.5 | $10 | $0.25 |
| Aug 2024 | $12.5 | $10 | $0.25 |
| Nov 2024 | $8.75 | $2.5 | $0.15 |
| Feb 2025 | $8.75 | $2.5 | $0.15 |
| May 2025 | $8.75 | $2.5 | $0.15 |
| Jun 2025 | $2.5 | $1.25 | $0.15 |
| Aug 2025 | $2.5 | $1.25 | $0.15 |
| Nov 2025 | $1.25 | $1.25 | $0.05 |
| Dec 2025 | $2 | $1.25 | $0.05 |
| Feb 2026 | $2 | $1.25 | $0.05 |
| May 2026 | $5 | $2 | $0.05 |
| Aug 2026 | $5 | $2 | $0.05 |
| Oct 2026 | $10 | $2 | $0.10 |
| Nov 2026 | $10 | $2 | $0.10 |
| May 2027, trend | $8.91 | $1.36 | $0.079 |
| Nov 2027, trend | $7.93 | $0.93 | $0.063 |
| Nov 2028, trend | $6.29 | $0.43 | $0.04 |
| Nov 2030, trend | $3.93 | $0.091 | $0.016 |
Source: live and archived API pricing pages of OpenAI, Anthropic, Google (via Vertex AI); SaaS Pricing Tracker analysis. See the full AI model API price history.
The long-run direction remains strongly deflationary: API inference prices are a fraction of their early-2020s levels. But the market has become less uniform. Cheap capability continues to commoditize while suppliers increasingly charge a premium for their newest frontier systems.
This makes a single question, “are AI prices going down?”, less useful. The more important questions are how quickly a fixed level of capability becomes cheaper, how much the newest available capability costs, and how much inference a useful workflow consumes.
Exhibit 3
The same level of AI capability is getting cheaper far faster than token prices suggest
External evidence on price/performance improvement; measures are not directly interchangeable
External research
~13× / year
historical decline in the cost of a fixed measured level of AI performance since 2023
Epoch AI, across five benchmark families. Historical result, not a prediction that the rate continues indefinitely.
Historical empirical, Epoch AI
How many times cheaper a year, log scale
Cost of reaching a fixed level of measured performance, five benchmark families
~13×~47% cheaper per quarter; ~13× cheaper per year since 2023
Fixed capability right after it first appears at the frontier
~75×~66% cheaper per quarter; ~75× per year for performance that has just debuted as state of the art
Same capability about two years after it first reached the frontier
~4.7×~32% cheaper per quarter; ~4.7× per year, half the pace of newly debuted performance
External forecasts, Gartner
Provider inference cost for a comparable 1-trillion-parameter LLM
More than 90% lower in 2030 than in 2025
Inference cost per agentic workflow
More than 5× higher through 2028
Show the data as a table
| Source | Date | What it measures | Result | Type |
|---|---|---|---|---|
| Epoch AI | Sep 2026 | Cost of reaching a fixed level of measured performance, five benchmark families | ~47% cheaper per quarter; ~13× cheaper per year since 2023 | Historical empirical |
| Epoch AI | Sep 2026 | Fixed capability right after it first appears at the frontier | ~66% cheaper per quarter; ~75× per year for performance that has just debuted as state of the art | Historical empirical |
| Epoch AI | Sep 2026 | Same capability about two years after it first reached the frontier | ~32% cheaper per quarter; ~4.7× per year, half the pace of newly debuted performance | Historical empirical |
| Gartner | Mar 2026 | Provider inference cost for a comparable 1-trillion-parameter LLM | More than 90% lower in 2030 than in 2025 | External forecast |
| Gartner | Aug 2026 | Inference cost per agentic workflow | More than 5× higher through 2028 | External forecast |
The distinction between list price and capability-adjusted cost is central to AI economics. Epoch AI estimates that, since 2023, the cost of obtaining a fixed level of measured AI performance has fallen about 47% per quarter, or roughly 13× per year.
The decline is not constant. In Epoch’s data, a newly reached frontier performance level becomes cheaper particularly quickly, then the rate moderates as that capability ages. This is one reason our 2030 scenario does not mechanically extrapolate 13× annual improvement.
Gartner’s forecasts illustrate the other half of the problem: the provider cost of running a roughly comparable model can fall by more than 90%, while the cost of a complete agentic workflow still rises because the workflow consumes vastly more inference.
Exhibit 4
We expect 2026-level AI capability to cost roughly 1% as much by 2030
AI Price Index scenarios; cost index, October 2026 = 100
AI Price Index scenario, not external consensus
| Year | Slower progress | Central | Rapid progress |
|---|---|---|---|
| 2026 | 100 | 100 | 100 |
| 2028 | 25 (4× cheaper) | 4 (25× cheaper) | 1 (100× cheaper) |
| 2030 | 10 (10× cheaper) | 1 (100× cheaper) | 0.1 (1,000× cheaper) |
Source: AI Price Index scenarios, informed by Epoch AI, The plunging price of thought and the exhibits above.
Our central case assumes dramatically slower improvement than a mechanical continuation of the recent ~13× annual historical rate. That reflects several reasons for caution: benchmark improvements do not transfer perfectly to production, the decline tends to slow as capabilities age, latency and reliability requirements matter, and providers may retain part of the efficiency gain through margins or spend it on more capable systems.
The result is still profound: our central scenario has the cost of reproducing approximately 2026-level AI capability falling about 25× by 2028 and 100× by 2030.
Exhibit 5
One dollar buys far more AI capability than token prices alone reveal
Artificial Analysis Intelligence Index (higher is better) against input price, USD per million tokens, log scale
Observed list prices
The blue line is the efficient frontier: the best score available at each price. The dash-dot line moves it by the central AI Price Index scenario for 2028.
Show the data as a table
| Model | Input price | Score |
|---|---|---|
| Opus 5.5 | $4 | 58 |
| Sonnet 5.5 | $2 | 56 |
| Fable 5.1 | $10 | 53 |
| gpt-6-astra | $10 | 53 |
| gpt-6.1-sol | $2 | 52 |
| grok-4.7 | $2 | 46 |
| gpt-5.6-terra | $2 | 42 |
| Gemini 3.8 Flash | $0.75 | 41 |
| deepseek-flash | $0.30 | 39 |
| gpt-6-luna | $0.10 | 38 |
| deepseek-v4-pro | $1.32 | 36 |
| Gemini 3.1 Pro Preview | $2 | 30 |
| Gemini 3.5 Flash-Lite | $0.30 | 22 |
| Haiku 4.5 | $1 | 17 |
| Mistral Medium 3.5 | $1.50 | 14 |
| Nova 2 Pro | $1.38 | 14 |
| Nova 2 Lite | $0.30 | 13 |
| Mistral Small 4 | $0.15 | 11 |
| Mistral Large 3 | $0.50 | 9 |
| Ministral 3 8B | $0.15 | 5 |
On our own price lines, GPT-4 Turbo-level capability fell from $10 to $0.10 per million input tokens
Cheapest tracked model scoring at least GPT-4 Turbo (127.25) on the Epoch Capabilities Index
- GPT-4 TurboMay 2024$10
- Claude 3.5 SonnetJun 2024$3
- GPT-4oNov 2024$2.50
- Gemini 2.0 FlashMar 2025$0.15
- Gemini 2.5 FlashSep 2025$0.30
- GPT-5 nanoNov 2025$0.05
- GPT-5.6 LunaSep 2026$0.20
- GPT-6 LunaOct 2026$0.10
Small rises in between are tier swaps, when a vendor replaces its cheapest model with a better but pricier one. GPT-6 Luna is not scored by Epoch yet; it outscores Claude Haiku 4.5 on the Artificial Analysis index, so it clears the bar.
Source: Artificial Analysis Intelligence Index, read Oct 2026, highest reasoning setting per model; Epoch Capabilities Index, read Oct 2026; list prices from each AI model API pricing page.
Token price is only the retail rate card. A model can be dramatically better value if it achieves nearly the same useful capability with fewer tokens, less reasoning or a much lower unit price.
Over time, the AI Price Index should increasingly measure capability per dollar and cost per successful task alongside raw token pricing. Comparable standardized task-cost data does not yet go back far enough for a historical index, so this exhibit shows today’s cross section; the historical series will expand as comparable data accumulates.
Exhibit 6
Newer GPUs cost more per hour, even as the cost of useful compute falls
Standardized GPU rental benchmarks, USD per GPU-hour, snapshot of 3 October 2026
External research
Neocloud index
- A100$1.54
- MI300X$2.63
- H100$2.77
- H200$3.40
- B200$5.86
- B300$6.70
Hyperscaler index
- A100$3.76
- H100$7.20
Bars are standardized market benchmarks, not directly comparable cloud SKUs. See AI infrastructure pricing for individual providers.
Show the data as a table
| GPU | Segment | USD / GPU-hour |
|---|---|---|
| A100 | Neocloud | $1.54 |
| MI300X | Neocloud | $2.63 |
| H100 | Neocloud | $2.77 |
| H200 | Neocloud | $3.40 |
| B200 | Neocloud | $5.86 |
| B300 | Neocloud | $6.70 |
| A100 | Hyperscaler | $3.76 |
| H100 | Hyperscaler | $7.20 |
Source: Silicon Data, Silicon Index; snapshot of the consolidated index page; the index family covers 95% of neocloud GPU providers and all major hyperscalers, published daily.
The latest accelerator does not need to rent for less than its predecessor for AI economics to improve. A B200 or B300 can command a higher hourly price while producing substantially more useful inference during that hour.
The relevant economic variable is therefore increasingly useful compute per dollar, not simply dollars per GPU-hour.
Exhibit 7
Better hardware does not guarantee cheaper GPU-hours
Silicon Data neocloud rental indices, 4 May 2026 to 27 Jul 2026
External research
Show the data as a table
| GPU | 4 May 2026 | 27 Jul 2026 | Change |
|---|---|---|---|
| H100 | $2.57 | $2.75 | +7.0% |
| H200 | $2.70 | $3.09 | +14.4% |
During this twelve-week period in 2026, standardized H100 rental prices rose 7.0% and H200 prices rose 14.4%. This does not establish a permanent direction, but it demonstrates why technological progress cannot be translated mechanically into lower nominal GPU-hour prices.
Strong demand, supply composition and the value of newer memory-rich accelerators can offset depreciation at the rental-price layer.
Exhibit 8
The compute market expects older GPU rental rates to decline, but not collapse
Silicon Data neocloud spot versus 36-month term rates as of 19 July 2026
External research
Show the data as a table
| GPU | Spot USD/hour | 36M term USD/hour | Difference |
|---|---|---|---|
| A100 | $1.65 | $1.40 | -15% |
| H100 | $2.72 | $2.38 | -13% |
| B200 | $5.62 | $5.17 | -8% |
Source: Silicon Data, GPU futures; market snapshot as of 19 July 2026.
Forward pricing was in backwardation across A100, H100 and B200 in July 2026: the market expected lower future rental rates, but only by single- to mid-teen percentages over three years.
This is useful counterweight to simplistic assumptions that hardware progress should drive the nominal GPU-hour price toward zero.
Exhibit 9
Enterprise cloud users pay a large premium for the same class of accelerator
Hyperscaler index ÷ neocloud index, USD per GPU-hour
External research
~2.6×
H100 hyperscaler premium
- H100
- Neocloud $2.77
Hyperscaler $7.20 - 2.60×
- A100
- Neocloud $1.54
Hyperscaler $3.76 - 2.44×
Source: Silicon Data, Silicon Index; consolidated index, snapshot of 3 October 2026.
Customers do not buy silicon alone. Hyperscaler pricing also reflects availability, networking, managed infrastructure, geography, enterprise agreements, security, compliance and integration.
The same phenomenon appears at the model layer: the raw cost of computation can fall rapidly without every supplier passing the reduction directly into its headline price.
Exhibit 10
Useful compute gets cheaper much faster than the newest GPU-hour
AI Price Index scenario, 2026 = 100, log scale per panel
AI Price Index scenario, not external consensus
Nominal price of a latest-generation GPU-hour
Useful compute delivered per GPU-hour
Effective compute cost per fixed unit of work
Show the data as a table
| Metric | 2026 | 2028 | 2030 |
|---|---|---|---|
| Nominal price of a latest-generation GPU-hour | 100 | 90–130 | 80–140 |
| Useful compute delivered per GPU-hour | 100 | 400–1,000 | 1,500–5,000 |
| Effective compute cost per fixed unit of work | 100 | 10–30 | 2–10 |
Source: AI Price Index scenario, informed by Exhibits 6 to 8.
Our central view is not that the newest accelerator becomes dramatically cheaper in nominal dollars. It is that each accelerator-hour produces dramatically more useful work.
By 2030, we model effective compute cost for a fixed unit of work at roughly 2–10% of today’s level even if the nominal price of the newest GPU-hour remains around today’s order of magnitude.
Exhibit 11
AI infrastructure investment is moving from hundreds of billions to trillions
Distinct estimates with different scopes; not additive
External research
Annual planned capex
~$800B
2026
Five major hyperscalers' planned capex, McKinsey
~$1.0T
2027
Five major hyperscalers' planned capex, McKinsey
Cumulative through 2030
Goldman Sachs Research
Capex of the four largest hyperscalers (Alphabet, Amazon, Meta, Microsoft), 2025–2030
~$5.3TMcKinsey
Global data-center physical infrastructure, excluding IT hardware
>$1.7TMcKinsey
Total global data-center capital requirement
~$7T
Each bar has its own scope. Do not add them together.
Show the data as a table
| Source | Scope | Period | Value |
|---|---|---|---|
| McKinsey | Five major hyperscalers' planned capex | 2026 | ~$800B |
| McKinsey | Five major hyperscalers' planned capex | 2027 | ~$1.0T |
| Goldman Sachs Research | Capex of the four largest hyperscalers (Alphabet, Amazon, Meta, Microsoft), 2025–2030 | Through 2030, cumulative | ~$5.3T |
| McKinsey | Global data-center physical infrastructure, excluding IT hardware | Through 2030, cumulative | >$1.7T |
| McKinsey | Total global data-center capital requirement | Through 2030, cumulative | ~$7T |
Source: McKinsey, The $2.3 trillion horizon; Goldman Sachs Research; McKinsey, The capex crucible; McKinsey, Scaling bigger, faster, cheaper data centers.
The scale of the buildout is a central part of the price story. Five major hyperscalers have announced about $800 billion of capex for 2026 and $1 trillion for 2027, according to McKinsey, with the vast majority directed toward AI infrastructure.
Goldman Sachs Research estimates the four largest hyperscalers will spend roughly $5.3 trillion on AI and data centers through 2030. McKinsey separately estimates around $7 trillion of total global data-center capital requirements through 2030, including $1.7–1.9 trillion of construction costs.
These estimates overlap and use different scopes. They should be read as evidence of the scale of the investment cycle, not summed into a single total.
Exhibit 12
Data-center electricity demand roughly doubles by 2030 while AI-focused consumption triples
IEA central projection, global data-center electricity consumption, TWh
External research
~3×
AI-focused data-center electricity, 2025 to 2030
Shown as a multiple: the IEA summary gives the growth factor, not a comparable absolute TWh value.
~3%
of global electricity demand from data centers in 2030
Show the data as a table
| Year | All data centers, TWh | AI-focused, index |
|---|---|---|
| 2025 | 485 | 100 |
| 2030 | 950 | ~300 |
The IEA’s updated central projection has global data-center electricity consumption rising from 485 TWh in 2025 to 950 TWh in 2030, around 3% of global electricity demand. Electricity use by AI-focused data centers grows considerably faster, roughly tripling over the same period.
Power is not a permanent token-price floor: model and hardware efficiency can improve faster than electricity costs. But grid connections, generation, transformers, cooling and construction constrain how quickly additional compute capacity can be deployed.
Exhibit 13
AI inference demand may already be growing faster than supply
Current external estimates are highly uncertain
External research
Current global inference demand at current prices, tokens per second, log scale
Current growth proxy: ~10× / year
Source: Epoch AI, Is a compute crunch coming?; Epoch calls these estimates highly uncertain.
Epoch AI estimates current inference demand at roughly 200 million to 4 billion tokens per second and cites proxy data suggesting growth around 10× per year. The range is intentionally wide because global token consumption, model mix and private inference are poorly observed.
We therefore use this as evidence of extraordinary current demand growth, not as a rate to extrapolate mechanically through 2030.
Exhibit 14
Our central case assumes inference demand grows faster than available capacity
AI Price Index scenario; 2026 = 100; log scale
AI Price Index scenario, not external consensus
2030 scenario ranges, multiple of 2026 level
Show the data as a table
| Year | Demand index | Capacity index |
|---|---|---|
| 2026 | 100 | 100 |
| 2027 | 500 | 300 |
| 2028 | 2,000 | 800 |
| 2029 | 6,000 | 1,800 |
| 2030 | 15,000 | 4,000 |
Source: AI Price Index scenario, informed by Epoch AI, Is a compute crunch coming? and Exhibits 11 and 12.
Current usage growth is too extreme and too uncertain to extrapolate directly. Our scenario therefore assumes substantial deceleration in both demand and capacity growth.
Even after that deceleration, the central case has inference use rising about 150× from 2026 to 2030 versus about 40× growth in available capacity. This is the mechanism behind our view that frontier compute can remain scarce while older capability becomes extremely cheap.
Exhibit 15
Cheaper chips do not translate one-for-one into cheaper AI tasks
The layers between an accelerator-hour and a finished task
Hardware and accelerator economics
Datacenter, electricity and networking
Inference software and utilization
batching · quantization · kernels · speculative decoding · parallelism
Model efficiency
Tokens and compute required by the workload
context · reasoning · output · retries · multimodal inputs
Caching, routing and tools
Provider margin, scarcity, service tier and strategic pricing
Cost per successful task
Cost per successful task = (input + cache reads + cache writes + reasoning/output + tools + retries + failed attempts) / accepted completed tasks
A token is an accounting unit, not the final economic output. Two models with identical token prices can have very different task economics if one requires fewer tokens, less reasoning, fewer retries or can reuse cached context.
For production AI, cost per successful task increasingly matters more than cost per input token.
Exhibit 16
AI gets cheaper, so we use vastly more of it
Measured external evidence and clearly labelled illustration
External forecast, Gartner, August 2026
AI inference costs per agentic workflow forecast to increase more than 5× through 2028.
Gartner also states that routing a task to an agentic reasoning model can cost a provider at least five times as much as a basic chatbot interaction, often more as complexity rises.
Illustrative economics, not a measured forecast
| Workload | Relative unit cost | Inference consumed | Workflow cost |
|---|---|---|---|
| Simple 2026 AI task | 100 | 1 | 100 |
| Same capability in 2030 | 1 | 1 | 1 |
| Frontier 2030 agent | 10–20 | 50–200+ | potentially >100 |
Relative values, simple 2026 task = 100 for unit and workflow cost. The multiplication is illustrative, not a precise forecast.
Intelligence becomes cheaper. We spend the savings buying more intelligence.
Source: Gartner, August 2026.
Falling unit costs unlock applications that would previously have been uneconomic. Reasoning systems spend more tokens thinking; agents repeatedly call models and tools; longer contexts keep more information active; and multimodal systems process richer inputs.
The result is a classic demand response: the price of a unit can collapse while expenditure per advanced workflow, and total expenditure across the economy, rises.
Exhibit 17
Competition can make today’s AI price lower than its sustainable economic cost
Observed API prices reflect strategy as well as technology
Conceptual bridge, not to scale; no values are estimated
- Underlying inference cost
- +Long-run required margin
- =Sustainable economic price
- −Strategic subsidy / competitive pricing
- =Observed customer price
Observed AI price = underlying cost + sustainable margin − strategic subsidy
Underlying cost
accelerator depreciation or rental · electricity · networking · data-center cost · utilization · inference software · model-specific compute requirements
Sustainable margin
R&D · training new models · operations · sales · capital costs · return on invested capital
Strategic subsidy
venture or equity capital · strategic investor capital · cloud credits · hyperscaler economics · customer-acquisition spending · ecosystem lock-in · developer adoption · cross-subsidy from another product · introductory, free-tier and promotional pricing
AI providers are not necessarily optimizing current-period inference profitability. Frontier laboratories and hyperscalers are simultaneously competing for developers, enterprise adoption, consumer market share and control of the future AI platform layer.
That creates an important forecasting risk: a very low API price today may reflect genuine technological deflation, strategic subsidy, or both. We therefore do not assume that every observed decline in market price represents an equivalent decline in sustainable cost.
Provider unit economics are mostly private. We treat subsidy as a market-structure risk and a forecasting variable, not as a known fact about any particular API, and we do not claim that any provider sells API tokens below cost.
Exhibit 18
The same technology can produce three different price paths, depending on market structure
AI Price Index market-structure scenarios; qualitative, no probabilities assigned
AI Price Index scenario, not external consensus
Scenario A
Subsidy persists
Competition remains extremely aggressive. Capital remains abundant. Providers keep prioritizing adoption and ecosystem control.
- Commodity API prices fall extremely quickly
- Prices may stay close to marginal cost
- Free and cheap tiers remain common
- API price declines can exceed underlying compute-cost declines
Scenario B, central case
Rational competition
Inference costs decline rapidly and competition remains strong, but providers nevertheless demand positive unit economics.
- Commodity prices decline substantially
- Frontier pricing remains premium
- Margins normalize gradually
- Observed price declines roughly track, but do not fully capture, technology improvements
Scenario C
Subsidy unwinds
Capital discipline increases or market concentration rises. Providers seek substantially higher margins.
- Some API prices can rise even while inference costs fall
- Free tiers shrink
- Discounts become less aggressive
- High-usage plans move toward metered pricing
- The gap between technological cost decline and customer price decline becomes very visible
Scenario B, rational competition, is the AI Price Index central case and underlies the central values in Exhibit 20. Scenarios A and C sit behind the lower and upper ends of its ranges.
The scenarios differ in the subsidy term, not in technology. Inference costs fall rapidly in all three. What changes is how much of that decline customers see, and when.
Exhibit 19
External research points to cheaper unit inference and a vastly larger AI infrastructure base
Evidence map; confidence labels are qualitative editorial judgments, not probabilities
External research
| Topic | External evidence | Horizon | Confidence |
|---|---|---|---|
| Fixed measured capability cost | Epoch: ~13× cheaper a year since 2023Epoch AI, The plunging price of thought | Historical | High for studied benchmarks |
| Comparable large-model provider inference cost | Gartner: >90% lowerGartner, March 2026 | 2030 vs 2025 | Medium |
| Agentic workflow inference cost | Gartner: >5× higherGartner, August 2026 | Through 2028 | Medium |
| Major hyperscaler capex | McKinsey: ~$800B in 2026, ~$1T in 2027McKinsey, The $2.3 trillion horizon | 2026–27 | Medium-high |
| Four largest hyperscalers, AI and data-center capex | Goldman Sachs: ~$5.3T, 2025–2030Goldman Sachs Research | Through 2030 | Medium |
| Overall data-center electricity | IEA: 485 → 950 TWhIEA, Energy and AI, executive summary | 2025–30 | Medium-high |
| AI-focused data-center electricity | IEA: ~3×IEA, Energy and AI, executive summary | 2025–30 | Medium-high |
| Current inference demand | Epoch: ~0.2–4B tokens a secondEpoch AI, Is a compute crunch coming? | Current | Low-medium |
| Current demand-growth proxy | Epoch: ~10× a yearEpoch AI, Is a compute crunch coming? | Current | Low |
| Below-cost AI access | OpenAI CEO: losing money on the $200-a-month ChatGPT Pro consumer planFortune, Sam Altman says OpenAI is losing money on ChatGPT Pro, January 2025 | Jan 2025 | High for that plan; says nothing about API margins |
The external evidence is unusually consistent on direction but not on exact magnitude. Equivalent-model inference becomes much cheaper, capital deployment and electricity demand rise sharply, and AI workloads become substantially more inference-intensive.
There is no rigorous “consensus token price for 2030” because these sources forecast different layers of the stack. The next section is therefore explicitly our synthesis rather than an average of incompatible forecasts.
Exhibit 20
Commodity AI keeps getting cheaper; frontier pricing remains in roughly today’s order of magnitude
AI Price Index market-price scenarios; USD per million uncached input tokens, log scale
AI Price Index scenario, not external consensus
A major uncertainty is how long AI providers are willing to subsidize inference. Current market prices are set during an unusually intense period of platform competition and capital investment. If suppliers eventually optimize for sustainable margins rather than market share, customer prices could fall materially more slowly than underlying inference costs, or temporarily rise even as the technology becomes cheaper.
Future market price = future cost-to-serve + future sustainable margin − future strategic subsidy
Median frontier
Cheapest frontier
Commodity / budget
Cost of 2026-level capability, index, October 2026 = 100
Show the data as a table
| Metric | Oct 2026 baseline | 2028 central | 2028 range | 2030 central | 2030 range |
|---|---|---|---|---|---|
| Median frontier | $10 | $10 | $5–$20 | $10 | $3–$30 |
| Cheapest frontier | $2 | $1.50 | $0.50–$4 | $1 | $0.20–$5 |
| Commodity / budget | $0.10 | $0.05 | $0.01–$0.20 | $0.02 | $0.003–$0.20 |
| Cost of 2026 capability, index | 100 | 4 | 1–25 | 1 | 0.1–10 |
The upper end of each range incorporates
- scarcity
- provider margin normalization
- reduced subsidies
- continued premium frontier positioning
The lower end incorporates
- extremely strong competition
- continuing strategic subsidy
- rapid efficiency gains
- open-model pressure
The future value of the subsidy term is unknown. Pricing normalization risk is one reason the ranges stay broad.
Source: AI Price Index scenarios; baseline from Exhibit 1.
We do not expect the price of frontier intelligence to follow the same curve as the price of reproducing today’s capability.
Our central 2030 case keeps the median frontier API in roughly today’s order of magnitude, around $10 per million uncached input tokens, while commodity inference falls toward roughly $0.01–0.03. Meanwhile, obtaining approximately 2026-level capability falls to around 1% of today’s cost.
The frontier price is not a hard floor. It is our market-structure view: suppliers can spend efficiency gains on larger models, longer reasoning, higher reliability, multimodality and more compute rather than passing every improvement through as a lower token rate.
We also do not forecast by applying a compute-cost decline to today’s list price. A 90% fall in the cost of compute does not imply a 90% lower API price when part of today’s price may be a strategic subsidy that later normalizes. The central values correspond to Scenario B, rational competition, in Exhibit 18.
Exhibit 21
AI becomes cheaper per unit of intelligence, and vastly larger as an industry
2030 synthesis
What gets cheaper
- 2026-level AI capability
- ~100× cheaper
- AI Price Index central scenario by 2030.
- Comparable-model provider inference
- >90% cheaper
- Gartner forecast, 2030 vs 2025.
- Commodity API input
- ~$0.01–0.03 / 1M tokens
- AI Price Index central market-price view.
- Effective compute for fixed work
- ~10–50× better
- AI Price Index scenario order of magnitude.
What gets bigger or stays expensive
- AI inference usage
- ~150×
- AI Price Index central 2026 to 2030 scenario.
- Inference capacity
- ~40×
- AI Price Index central scenario.
- Overall data-center electricity
- 485 → 950 TWh
- IEA, 2025 to 2030.
- AI-focused electricity
- ~3×
- IEA, 2025 to 2030.
- Four largest hyperscalers, AI and data-center capex
- ~$5.3T
- Goldman Sachs Research, 2025 to 2030.
- Median frontier API price
- ~$10 / 1M input
- AI Price Index central 2030 view.
The apparent contradiction is the point. Technological deflation makes each unit of useful intelligence much cheaper, which expands the set of economical applications. The additional demand then drives more infrastructure, more tokens, more reasoning and more aggregate spending.
- Technological cost
- Strongly downward
- Sustainable economic price
- Downward, but probably much more slowly
- Observed market price
- Can temporarily sit above or below sustainable economics
The lowest market price is therefore not necessarily the best predictor of the long-run cost of AI. During a platform land-grab, price can be a strategic weapon.
AI Price Index view
Our view: cheap intelligence, expensive frontier
- 1
Commodity intelligence trends toward negligible cost
Capabilities that require expensive frontier systems today should migrate into smaller, cheaper and more specialized models. The economic value of yesterday’s frontier is therefore likely to commoditize quickly.
- 2
The frontier does not become free
The frontier itself keeps moving. Providers can spend efficiency gains on larger models, longer reasoning, multimodal processing, latency and reliability rather than simply lowering the list price.
- 3
Demand is likely to grow faster than prices fall
Lower prices unlock new workloads. Agents, background automation, generated media, coding systems and machine-to-machine inference can consume orders of magnitude more computation than today’s chat interactions.
- 4
Compute per dollar improves much faster than GPU-hour prices fall
The latest accelerator can remain expensive per hour while becoming dramatically cheaper per useful unit of computation.
- 5
Token price becomes a progressively weaker economic metric
Capability per dollar and cost per successful task should increasingly be tracked alongside headline input and output token rates.
- 6
AI becomes simultaneously cheap and expensive
2026-level intelligence becomes commodity-priced. The 2030 frontier remains premium-priced. Aggregate spending on AI becomes enormous.
- 7
Today’s price may not be the equilibrium price
AI is still in a platform land-grab. Providers have strong incentives to acquire developers, users and enterprise workloads before the market structure stabilizes. Some products can therefore be priced at margins that would not make sense in a mature market. We expect technological inference costs to continue falling rapidly, but customer prices may not fall at the same rate if today’s competitive subsidies eventually normalize.
By 2030, today’s intelligence may cost around one-hundredth as much while the world spends vastly more producing and consuming intelligence overall.
Quantum computing could eventually change the cost curve, but it is not in our central forecast
0%
quantum contribution included in the AI Price Index central inference-cost forecast through 2030
Direct
Quantum algorithms eventually accelerate economically relevant parts of AI computation.
Indirect
Quantum simulation contributes to semiconductor and material discovery, chemistry, energy technologies, cooling and batteries.
AI → quantum
AI improves calibration, control, error correction, algorithm discovery and system optimization.
The AI Price Index will incorporate quantum into its quantitative AI cost model only after a quantum or hybrid system demonstrates lower end-to-end economic cost for a commercially relevant workload than the contemporary classical alternative.
Questions people ask
Are AI API prices going down?
Yes, but the answer depends on what is being measured. The cheapest useful models have become dramatically less expensive, while premium frontier pricing has recently diverged upward. More importantly, research suggests the cost of obtaining a fixed level of AI capability is falling much faster than headline token prices.
How much will AI cost in 2030?
Our central scenario is that by 2030 the cost of obtaining approximately 2026-level AI capability falls by around 99%, while frontier API prices remain in roughly the same order of magnitude as today. Commodity inference could fall toward roughly $0.01–0.03 per million input tokens, but increasingly complex reasoning and agentic workloads mean total AI spending is likely to rise rather than fall.
Why can AI spending rise if AI gets cheaper?
Lower unit costs make far more AI usage economical. Agents, reasoning systems, coding tools and background automation consume many more inference steps than a simple chatbot interaction. Falling costs therefore expand demand and can increase total spending, a version of the Jevons paradox.
Will GPU prices fall?
Effective compute costs should continue falling rapidly, but the rental price of the newest GPU does not have to. New accelerators can command premium hourly rates while delivering much more inference per hour, making useful compute cheaper even when the hardware-hour remains expensive.
Are current AI API prices sustainable?
We do not know for every provider. Model-serving costs, utilization and negotiated compute prices are mostly private. In a rapidly growing platform market, suppliers may intentionally accept unusually low margins to gain adoption, and some AI access products have demonstrably been priced below realized usage cost. For that reason, the AI Price Index treats market price and underlying economic cost as separate concepts.
Could AI prices rise even if compute gets cheaper?
Yes. Prices can rise if strategic subsidies disappear, providers seek higher margins, competition declines, demand outruns capacity, or a new frontier model consumes much more compute. Falling technological cost creates downward pressure on price, but it does not mechanically determine the customer price.
AI capability is likely to get much cheaper, but AI as an industry is not going to get smaller. Falling unit costs will drive more consumption, larger models, longer reasoning and more autonomous workloads. By 2030, today’s intelligence may cost around one-hundredth as much, while the industry spends vastly more delivering intelligence overall.
AI prices are being driven by two extraordinary forces at the same time: technological deflation and a platform land-grab. Technology makes intelligence cheaper to produce. Competition can make it even cheaper to buy. But only the first of those effects is guaranteed to survive a normalization of market structure.
Methodology
Frontier
Each model maker contributes one frontier line: the input price of its top general-purpose model on its own API pricing page (OpenAI: Flagship model; Anthropic: Top model; Google: Gemini Pro tier; Mistral: Mistral Large 3; xAI: grok-4.6; DeepSeek: deepseek-v4-pro). The line follows whichever model holds that position at each date. The median and cheapest frontier series are the median and minimum across makers in the index.
Budget
Each maker also contributes one budget line, the small, fast tier it lists on its own pricing page (OpenAI: Cheapest model; Anthropic: Haiku tier; Google: Gemini Flash tier; Mistral: Mistral Small 4; xAI: grok-4.3; DeepSeek: deepseek-flash). The budget series is the minimum across makers.
Provider entry
A maker joins after a year of history. Resellers are left out, so a model is never counted twice; Google is read through Vertex AI, which matches the Gemini API list price. The historical basket is not rewritten when new makers join.
Prices and dates
Prices are standard, uncached list prices per million input tokens. Each observation is dated by the live or archived pricing page that shows it, so a price change happened on or before that date. Capture dates are not yet separated from effective dates; where they differ, the index lags the real change.
Market price versus economic cost
The AI Price Index primarily observes customer-facing market prices. These should not be interpreted as direct measurements of providers’ underlying inference costs or sustainable long-run prices. Providers can temporarily price above cost because frontier capacity is scarce, or below sustainable economics because market share and ecosystem adoption are strategically valuable.
Where reliable evidence exists, we separately identify promotional pricing, unusual discounting and documented below-cost access. We do not estimate hidden provider margins without sufficient evidence.
Three kinds of price decline
- Technology-driven deflation. Actual serving cost falls because of better hardware, better models, inference optimization, higher utilization, caching or quantization. This is the most structurally durable type of decline.
- Competitive margin compression. Providers can serve profitably but accept lower margins because competition is intense. This can persist for a long time, but it is not the same as a fall in technological cost.
- Strategic subsidy or loss-leading pricing. Price sits below the level that would support sustainable standalone economics. This is the least safe kind of decline to extrapolate indefinitely.
The historical market-price series in Exhibit 1 contains all three effects. The index does not separate them observation by observation.
Four price curves
- 1. Frontier access price
- What customers pay for the newest capability. Observed, Exhibit 1.
- 2. Commodity market price
- What customers pay for inexpensive inference. Observed, Exhibit 1.
- 3. Fixed-capability cost
- What it takes to reproduce a constant level of capability. External evidence, Exhibit 3.
- 4. Sustainable economic price
- The price that covers serving cost, capital requirements and a sustainable provider margin, without assuming indefinite strategic subsidy. Only partially observable; not estimated.
The fourth curve is not fabricated. It explains why today’s very low prices may not be a reliable long-run floor.
Price sustainability
Prices are classified qualitatively, never as a “true cost per token”. A company’s operating losses include training, R&D, employees, capex and customer acquisition, so they do not prove negative gross margin on a particular API request, and are never treated as API-level subsidy evidence.
- Unit economics unknown
- The default. Serving costs, utilization and negotiated compute prices are private, so no judgment is made.
- Apparently sustainable
- Direct evidence suggests the price covers serving cost with a positive margin.
- Uncertain
- Indirect evidence exists but points in no clear direction.
- Potentially subsidized
- A credible source indicates the price, or the product it belongs to, sits below realized cost.
- Explicitly promotional
- The supplier itself labels the price as introductory, temporary or discounted.
Status taxonomy
- Observed
- A price verified directly for that date. Every point in Exhibit 1 is observed.
- Reconstructed
- A historical price whose effective period is reconstructed from a dated primary announcement or archive.
- Announced
- A supplier has announced the future change, but it is not yet effective.
- Statistical trend
- A mechanical extrapolation of recent historical prices.
- External forecast
- A forecast published by an outside source.
- AI Price Index scenario
- Our explicit analytical assumption or forecast.
Sources and evidence
- AI price history: current LLM API prices, AI model API price history, and the live and archived pricing pages linked from each model maker’s page.
- Epoch AI, The plunging price of thought
- Epoch AI, Is a compute crunch coming?
- Gartner, March 2026
- Gartner, August 2026
- Silicon Data, Silicon Index
- Silicon Data, H200 vs H100 rental prices, May to July 2026
- Silicon Data, GPU futures
- McKinsey, The $2.3 trillion horizon
- Goldman Sachs Research
- McKinsey, The capex crucible
- McKinsey, Scaling bigger, faster, cheaper data centers
- IEA, Energy and AI, executive summary
- Fortune, Sam Altman says OpenAI is losing money on ChatGPT Pro, January 2025
- Artificial Analysis Intelligence Index and Epoch Capabilities Index, read Oct 2026.
Raw data and further reading
What each model maker in the basket charges today, USD per million tokens (input / output).
| Maker | Frontier | Budget | In index | Unit economics |
|---|---|---|---|---|
| OpenAI | $10 / $50Flagship model | $0.10 / $0.50Cheapest model | since 2021 | Unit economics unknown |
| Anthropic | $10 / $50Top model | $1 / $5Haiku tier | since 2024 | Unit economics unknown |
| $2 / $12Gemini Pro tier | $0.75 / $3.75Gemini Flash tier | since 2025 | Unit economics unknown | |
| Mistral | $0.50 / $1.50Mistral Large 3 | $0.15 / $0.60Mistral Small 4 | from 2027 | Unit economics unknown |
| xAI | $2 / $6grok-4.6 | $1.25 / $2.50grok-4.3 | from 2027 | Unit economics unknown |
| DeepSeek | $1.32 / $3.96deepseek-v4-pro | $0.30 / $1.20deepseek-flash | from 2027 | Unit economics unknown |