The price of a token
Considerations on token prices, and where token market share is going. Companion to Who serves the tokens? Editorial analysis, current as of early 2026.
The strangest market in technology
Tokens are sold in a market where the unit price collapses by roughly 90% a year while total spending grows by hundreds of percent — both at once, for years. GPT-4 launched in March 2023 at about $30 per million input tokens and $60 per million output; two years later, GPT-4-class quality could be had for well under a dollar, DeepSeek and Qwen were serving frontier-adjacent capability for cents, and yet aggregate inference spend — and the datacenter buildout behind it — kept setting records. Understanding how both curves can be true is most of understanding the AI business.
Six dynamics
LLMflation
The price of a constant level of capability falls roughly an order of magnitude per year: GPT-4-class answers cost ~$30–60 per million tokens in 2023 and pennies by 2025, as smaller models catch up and inference stacks improve. Deflation this fast has almost no precedent outside transistors.
Price per token vs price per task
Reasoning models flipped the metric: they charge less per token but think in thousands of hidden tokens per answer. What the buyer actually pays for is a completed task — and per task, frontier spend is falling far more slowly than the per-token sticker suggests.
The Jevons engine
Every price collapse multiplies consumption: cheaper tokens make agents viable, and one agentic workflow burns 100× the tokens of a chat. Unit price down ten-fold, volume up hundred-fold — total spend rises. Efficiency is the demand engine, exactly as Jevons observed with coal.
The barbell market
Commodity intelligence races to zero — open-weights and Chinese models set a world floor price near the cost of electrons. Frontier capability holds premium pricing because for high-stakes work (code that ships, legal drafts, agents with credentials) quality is the cost, not the tokens.
Tokens ≠ revenue
The volume leaders and the revenue leaders are different companies: consumer free tiers generate oceans of unmonetized tokens, while enterprise APIs monetize scarce, expensive ones — how Anthropic can lead enterprise API spend while serving a fraction of the industry's tokens.
Distribution vs trust
At the commodity end, distribution wins: the default assistant in your search bar, phone or office suite captures volume regardless of benchmarks. At the premium end, trust wins: reliability, safety and agentic competence are what enterprises actually renew on.
Where token share is going
- The floor rises, the frontier holds. Open-weights and Chinese models keep pushing commodity token prices toward electricity cost, absorbing the price-sensitive middle of the market; frontier vendors retreat upward into reasoning, agents and enterprise trust, where margins live. Expect share of tokens to keep diffusing while share of revenue concentrates.
- Consumer volume consolidates around distribution. Assistants embedded in search, phones and office suites will serve the majority of raw tokens, mostly free and increasingly ad- or bundle-monetized — a Google-shaped game in which model quality matters less than surface ownership.
- Enterprise share follows agents. The fastest-growing paid workload is agentic — coding above all — where tokens are burned in bulk and switching costs come from workflow integration, not chat habit. This is the segment where Anthropic overtook OpenAI in API spend in 2025, and where the next share battles will be decided.
- The metric will shift from tokens to tasks. Per-token pricing fits completion APIs, not agents that own outcomes. Expect outcome-adjacent pricing — per task, per seat-plus-usage, per resolved ticket — to spread, making token share an input metric rather than the scoreboard.
- Inference eats the capex. As reasoning and agents shift compute from training runs to serving, the economics increasingly resemble a utility: high fixed cost, collapsing marginal price, volume growth as the only way out. The winners will look less like software companies and more like power companies with exceptional margins at the top of the stack.