Inference Economics

In late July 2025 Anthropic told subscribers to its Claude Pro plan ($20 a month) and Max plans ($100 and $200 a month) that from 28 August it would add weekly usage limits to Claude Code, its coding agent. The company said some users were running the agent "continuously in the background, 24/7", and that the limits would affect fewer than 5% of subscribers (company statement, reported). Claude Code had been released as a research preview on 24 February 2025 and made generally available with the Claude 4 models on 22 May. By December 2025 Anthropic said it had reached a run rate of about $1 billion (company), and by February 2026 more than $2.5 billion (company claim, reported).

The limits were notable because of who imposed them. Anthropic owned the model and the product and bought the compute at cost, and it still could not offer unlimited agent use at a flat price. Six weeks earlier Cursor, which bought its intelligence from Anthropic and others, had faced the same problem from the other side: it replaced request-based limits on its $20 Pro plan with a usage pool billed at API rates in June 2025, and then apologised and refunded unexpected charges between 16 June and 4 July (company blog, reported). Over the same years the price of any fixed level of capability had fallen faster than almost any price in economic history. OpenAI's GPT-4 cost $30 per million input tokens and $60 per million output tokens at launch in March 2023; GPT-4o mini, which scored higher than the original GPT-4 on the MMLU benchmark, cost $0.15 and $0.60 in July 2024 (company price lists).

This chapter explains how both things can be true: why the price of a token falls so fast, why the cost of a task can rise anyway, and who keeps the difference.