The Token Price Collapse: Why AI Gets 10× Cheaper Every Year, and Who Pays for It
In November 2021 a GPT-3-level answer cost $60 per million tokens. Three years later the same capability cost six cents. Few products in history have cheapened this fast. Yet AI bills are rising, frontier prices are not falling, and the labs selling the tokens are fighting for margin. This guide untangles the three different "prices" of AI, where the savings come from, why cheaper tokens mean bigger bills, and who ends up paying.
From $60 to six cents
Andreessen Horowitz partner Guido Appenzeller picked a simple yardstick in November 2024: the cheapest model that scores 42 on the MMLU knowledge benchmark. When GPT-3 opened to the public in November 2021, it was the only model that could, at $60 per million tokens. Three years later the cheapest model to reach the same score, Meta's small Llama 3.2 3B served by Together.ai, cost $0.06: a thousand times less. He called it "LLMflation" and estimated that the cost of a given level of capability falls about 10× every year.
For comparison, Moore's law roughly doubled chip performance every two years. A tenfold fall in price every year is far faster. But the headline hides a twist: the price of the best models has not collapsed, and the amount people spend on AI keeps rising.
Three different prices of AI
Most confusion comes from mixing up three numbers. Click each.
Price per token vs price of capability vs cost per task
How fast is it falling? Four estimates
Every serious study agrees the price of a fixed capability is falling fast; they disagree on how fast, because they pick different benchmarks, thresholds and time windows.
Annual price decline for fixed capability, by study
Log scale; × per year. Where a study gives a range, the bar shows the central or typical figure and the label shows the range. Stanford's figure is converted from 280× over about two years.
- Epoch AI (March 2025): the price to reach a fixed benchmark score fell 9× to 900× per year depending on the level, with a median of 50×; using only data from 2024 onward, the median was 200×. Epoch noted the fastest drops were recent, so they may not last. A later Epoch estimate, reported secondhand, puts a fixed performance level about 47% cheaper per quarter (roughly 13× a year).
- Stanford AI Index 2025: GPT-3.5-level performance (about 64.8% on MMLU) fell from about $20 to $0.07 per million tokens between November 2022 and October 2024, achieved by Gemini 1.5 Flash-8B.
- MIT FutureTech, "The Price of Progress" (Gundlach and colleagues): a given benchmark level gets about 5–10× cheaper per year for frontier-quality models, and only around 3× per year of that is better algorithms; the rest is cheaper hardware and competition from open-weight models.
Two price stories: capability vs the frontier
Plot fixed-capability prices next to what the newest flagship costs and two stories appear. Old capabilities become nearly free; the frontier's price falls far more slowly and, in places, rises.
Price per million tokens, 2021 to 2026
Log scale. Fixed-capability lines join the two endpoints each study reports (the path between is not measured). Flagship points are OpenAI list input prices from secondary sources. Hover or tap a point.
The MIT paper's authors reportedly find that the price of running the very latest frontier models is rising, because models get bigger and reason for longer, even as yesterday's capability gets cheaper. Secondary coverage also reports OpenAI raising GPT-5.5's price. The token market is splitting in two: commodity intelligence heading toward zero, and frontier intelligence holding its price.
The October 2026 price ladder
List prices per million tokens span a factor of about 100 from the cheapest capable model to the most expensive flagship. Toggle input and output prices.
Selected API prices, US$ per million tokens
From price aggregators in October 2026; Gemini 4 Argon is an introductory price until 30 November 2026 ($4/$20 after). Check providers' pages for current prices.
Where the savings come from
No single breakthrough explains a tenfold yearly fall. Several independent improvements multiply together. Click a driver.
Six drivers of cheaper tokens
Figures are each source's own claims on its own models and hardware; they are not additive and cannot be compared directly. A 2026 paper, the "Inference Engineering Pareto Atlas", tries to put them on common axes.
Cheaper tokens, bigger bills
In 1865 the economist William Stanley Jevons noticed that more efficient steam engines made Britain burn more coal, not less, because cheaper energy opened new uses. AI tokens follow the same pattern.
Token volume is exploding
Secondary reports; Google's figures conflict between sources. Log scale.
- A natural experiment: when OpenRouter discounted OpenAI's GPT-5.6 Luna, its daily token volume jumped 13.8× over the pre-discount average (5.6× for GPT-5.6 Terra), and users largely kept using it after the promotion ended. OpenRouter cautions part of this came from a few very large accounts, and its own 2025 study found demand fairly price-insensitive at the macro level.
- Agents eat tokens: the FinOps Foundation reports reasoning and agentic workloads use 5 to 30 times more tokens per task than an equivalent chat exchange; AT&T's daily usage reportedly rose from about 8 billion to 27 billion tokens after deploying multi-agent systems.
- Hidden thinking: reasoning models generate internal "thinking" tokens you pay for but do not see; one analysis puts them at 1.2 to 9 times the visible output.
Cost per task calculator
The number that matters to a business is not the price per token but the cost per successful task. A cheap model that fails half the time and needs retries can cost more than an expensive one that succeeds first time.
What does one task really cost?
Token counts per task and success rates are illustrative assumptions, anchored to the reported 5–30× token gap between chat and agent work. Prices are the October 2026 list prices above. Failed attempts are retried until success, and each failure also costs the review or rework time you set.
Who pays? The squeeze on the labs
Falling prices are great for buyers. For the companies selling tokens, they collide with rising compute bills. Reported gross margins vary with source and method:
Reported AI lab gross margins and targets
Secondary reporting (The Information and analyst estimates); methods differ, especially on whether free users' compute is included.
Those margins are far below the roughly 75% Nvidia earns selling the chips (see Custom Silicon vs Nvidia). The labs' response is the same everywhere: cheaper hardware of their own, more efficient models, and pricing that charges more for the frontier and for priority. Our AI Factory calculator found a gigawatt site needs roughly $2.83 per million tokens just to break even on a six-year payback, under its illustrative assumptions. Budget models listed at $0.10 to $0.50 per million input tokens only work because they are small and run on heavily used, depreciated or highly efficient hardware.
DeepSeek and the price war
China's AI price war began when DeepSeek released its V2 model in May 2024 at prices far below Western rivals. After its R1 reasoning model shook markets in January 2025, DeepSeek cut prices further: from February 2025 it offered R1 75% cheaper and V3 50% cheaper during an off-peak window (16:30 to 00:30 GMT, daytime in Europe and the US). It ended the off-peak discount on 5 September 2025 in favour of flat rates of about $0.56 per million input tokens and $1.68 for output; one report says it later added a peak-hour surcharge in Beijing business hours. Its V4.1 Flash model, with open weights, lists at about $0.30 / $1.20 on OpenRouter.
Open weights matter as much as list prices: once a capable model is downloadable, dozens of hosting companies compete to serve it, pushing its price toward the cost of the hardware. MIT's analysis counts open-model competition as one of the main reasons fixed-capability prices fall.
Price tiers and discounts: a maturing market
| Mechanism | How it works | Example |
|---|---|---|
| Model tiers | The same lab sells small, mid and frontier models at very different prices | GPT-6 Luna $0.10 vs GPT-6 Astra $10 per million input tokens |
| Batch processing | Accept a delay (hours) and pay less | Claude Sonnet 5.5 batch: 50% off |
| Cached input | Re-used prompt prefixes are billed far cheaper | Cached input at a tenth to a twentieth of the normal input price (e.g. Sonnet 5.5 $0.20 vs $2; GPT-6.1 Sol $0.10 vs $2) |
| Long-context surcharge | Very long prompts cost more per token | GPT-6.1 Sol prompts over 272K tokens: $4 / $15 instead of $2 / $10 |
| Time of day | Cheaper when data centers are idle, dearer at peak | DeepSeek's off-peak discount (2025) and reported peak surcharge |
| Introductory pricing | Launch discounts to win share | Gemini 4 Argon $2 / $10 until 30 November 2026, then $4 / $20 |
These are the tools of mature utilities and airlines: charge by urgency, time and quality of service. Their spread across AI pricing in 2025 and 2026 is a sign tokens are becoming a commodity priced like electricity.
What it means
🛠️ For builders
Design for the price of next year, not today: a feature too expensive now may cost a tenth in 12 months. Measure cost per successful task, route easy work to cheap models and hard work to frontier ones, and use batch and cache discounts.💼 For buyers
Expect per-token prices to fall and total bills to rise as agents and reasoning spread. Budget by task volume, and watch hidden thinking tokens.📈 For investors
Commodity inference is a margin squeeze; profits accrue to those with the cheapest hardware (custom chips, efficient AI factories), the best frontier models, or distribution. Watch lab gross margins against their 70% targets.Reading the numbers: where sources disagree
| Claim | What the sources say |
|---|---|
| "Prices fall 10× a year" | a16z's 2021–2024 estimate. Epoch: 9–900× (median 50×, 200× using 2024+ data); MIT: 5–10× for frontier-level capability; a later Epoch estimate (secondhand): ~13×. |
| Frontier prices "rising 3–18× a year" | Attributed to the MIT paper's updated abstract by one summary; we could not confirm it in the abstract itself. |
| OpenAI's historic list prices | From secondary pricing pages; GPT-4 ($30/$60), GPT-4 Turbo ($10/$30), GPT-4o ($5/$15, then $2.50/$10), GPT-5 ($1.25/$10). |
| Google's monthly tokens | 480 trillion (May 2025) to ~3.2 quadrillion (May 2026) in one series; another source gives 1.3 quadrillion. Check Google's own figures. |
| OpenRouter volume | ~4.5 quadrillion tokens annual run rate (Aug 2026), doubling every 11 weeks, per secondary reporting. |
| Lab gross margins | OpenAI: ~40% (2024) to ~33% (early 2026) in one analysis vs a ~46% company projection for 2025. Anthropic: 2025 forecast cut to ~40% (~38% including free users). |
| Driver speed-ups (quantization 4–6×, batching 3–10×, speculative decoding 2–3×) | Practitioner and vendor claims on specific setups; not additive. |
| Price elasticity "10% cheaper → 12–18% more use" | Cited in one summary; original report not found. Not used. |
| Calculator inputs | Token counts and success rates are our illustrative assumptions. |
🔗 References and papers
- Studies of the price trend
- a16z: Welcome to LLMflation
- Epoch AI: LLM inference price trends · How has the cost of LLM inference changed?
- Gundlach et al., The Price of Progress (arXiv 2511.23455) · MIT FutureTech page
- VentureBeat on the Stanford AI Index 2025
- Inference economics of language models (arXiv 2506.04645)
- The Inference Engineering Pareto Atlas (arXiv 2609.17863)
- Volume, Jevons and cost per task
- OpenRouter: State of AI, 100T-token usage study
- OfficeChai: OpenRouter's GPT-5.6 discount experiment
- AI token spending in 2026: prices, volumes and what agents cost
- Benedict Evans: ways to think about token pricing
- Hexaware: the rate card stopped predicting the bill
- Prices and the price war
- BenchLM: LLM API pricing trends (Oct 2026) · Frontier model API price index
- Reuters via Business Standard: DeepSeek's off-peak discounts · CloudZero: DeepSeek pricing changes
- Margins
- Anthropic lowers its gross margin forecast (via The Information)
- AI Compute Stack on this site
- Part 2: Inside the AI Factory
- Part 4: The Memory Wall
- Part 5: Custom Silicon vs Nvidia
- Frontier Model Pricing