Tokens Get 13x Cheaper Every Year, Yet AI Bills Keep Growing: Breaking Down LLM Economics

The paradox seems illogical: tokens get an order of magnitude cheaper each year, yet companies' AI bills keep growing. The author of an analysis on Habr gathered data from Anthropic's prospectus (S-1), the Ramp spending index and Epoch AI research, and showed there is no contradiction. There are simply two different price curves.

Two Curves Instead of One

Epoch AI estimates that reaching a fixed level of quality has been getting roughly 13x cheaper per year since 2023. Extrapolated, over three-plus years that comes to on the order of ten thousand times (this is the analysis author's estimate, not a figure from the report). But companies don't pay for yesterday's level; they pay for the newest model. Its price stays in a corridor of $10–75 per million output tokens, and the same task consumes more tokens. In 2026 a direct price increase was added to this: new versions launched at higher prices than their predecessors.

The price list, meanwhile, reflects reality poorly. According to Ramp, American companies pay OpenAI and Anthropic an average of about $0.56 per million tokens. In agent scenarios, the bulk of volume is repeated reading of context from cache and cheap models. An example using Martin Alderson's methodology: a coding agent session with 10 million tokens from cache, 0.5 million input and 0.2 million output costs from 23 cents to $25. That is a two-order-of-magnitude gap, and the choice of model and caching architecture affects the budget more than any change to the price list.

Where the Provider's Money Goes

Anthropic's prospectus provided verifiable figures for the first time. Gross margin on serving requests is about 85% in the first half of 2026 (this is a calculation from revenue and cost of revenue; the company does not disclose it). Of every dollar of revenue, about 15 cents goes to inference, about 60 to R&D and training, about 23 to sales and about 6 to administration. About 80% of revenue comes from pay-per-token billing, the rest from subscriptions. A caveat: almost half of revenue flows through the AWS and Google Cloud marketplaces, and their commission is recorded in sales expenses, so the net margin is lower, but the order of magnitude is the same.

OpenAI presents a different picture: according to The Information and the WSJ, the operating loss in the first and second quarters of 2026 was $9.3 billion and $12.3 billion on revenue of $5.7 billion and $6.7 billion. Gross margin was about 39% in the first quarter, because of many free and cheap plans. These figures are estimates; the company does not disclose its financials.

The upshot: the price of a frontier model is determined not by the cost of a token, but by the need to recoup training of short-lived models, a premium for leadership, and capacity shortages. Notably, a month after the launch of Claude Fable 5, it accounted for only about 6% of the tokens purchased from Anthropic. It appears that willingness to pay for quality gains has a limit.

Subscriptions: Limits Instead of Price Hikes

A fixed subscription price works poorly when an agent autonomously burns tokens for tens of minutes. Providers respond not by raising prices but by cutting limits. In the author's model, where cost equals roughly 15% of the API price for the same volume, an average developer using Claude Code on the $100 plan brings in a margin of about $74. Even the 90th percentile stays in the black. Only intensive users are unprofitable: for extreme ones, the loss is estimated at about $530 per month. For a $200 plan portfolio with 15% heavy users, the margin is about 56%. This is a model with assumptions; some consumption profiles are taken as assumptions.

What to Do in Practice

First: count the cost of a task, not of a token. Look at the cache-read rate and the number of tokens per result. Second: route. Use the expensive model where the quality difference translates into money, and hand the rest to mid-tier flagships or cheap and open models. Third: build into your product the risk of changes to pricing and limits, since Anthropic's S-1 allows for a move to payment per outcome or per completed task. Fourth: don't expect the falling price of intelligence to lower your bill on its own, because agents' appetite is growing faster.