Cursor Pricing Backlash Exposes the Cost of AI Coding Tools

A feature sprint meets a billing shock

Through 2025, Cursor shipped features at a pace few developer tools could match. Then a pricing change undercut much of the goodwill that pace had earned. The AI code editor, built by Anysphere, spent the year adding autonomous agents and enterprise integrations. In mid-2025 it also rewrote the terms of its flagship subscription in a way many users found confusing and, for some, expensive.

What changed on June 16

The core facts are consistent across reporting. On June 16, 2025, Cursor reworked its $20-per-month Pro plan and introduced a $200-per-month Ultra tier marketed as offering 20 times Pro's usage.1 Under the old Pro arrangement, subscribers got 500 fast responses on premium models from OpenAI, Anthropic and Google, followed by unlimited responses at a slower speed.3 The new version replaced that with $20 worth of monthly usage billed at API rates, a shift Cursor described as moving from request limits to compute limits.13

The practical effect hit heavy users hardest. Some reported exhausting their allowance after only a handful of prompts, particularly when using Anthropic's newer Claude models, which are popular for coding.3 Others said they were billed extra charges they had not realized they were agreeing to once they went past the $20 threshold.3 Analytics India Magazine traced the visible surge in complaints to around July 4, when users began noticing their quotas draining far faster than expected.4 Its article carries a May date that does not match the July events it describes, so the July timeline from other outlets is the more reliable one.

"Unlimited" and the trust problem

The anger centered on one word. Many developers believed Pro was effectively unlimited, and a widely shared Reddit post argued that Cursor had "nuked" the plan and that "unlimited" had never really been unlimited.1 Users accused the company of a "rug pull," and reports described cancellations and public complaints spreading across social platforms.4 Accounts of a mass "exodus" rest mostly on that social-media evidence. No hard churn figures have been published, so the scale of the departures is hard to verify.

Cursor did not dispute that it mishandled the rollout. CEO Michael Truell apologized in a blog post, saying the company had not handled the change well and that its communication was neither clear enough nor expected by many users.3 Cursor also said it had "missed the mark" and would fully refund affected users.4

The economics underneath

The more revealing part of the story is why the change happened at all. Cursor's costs rise with usage. As power users relied more heavily on the tool, the cost of serving each of them grew.2 Model upgrades made that worse. Anthropic moved from Claude Sonnet 3.5 to 3.7 to Claude 4, and each release was more capable and more expensive per token.2 A flat $20 subscription with generous slow-mode access was hard to sustain against those costs.

The timing adds an irony. Cursor's 0.50 release in May 2025 was promoted partly for introducing unified request-based pricing, which was meant to replace the confusing token-based schemes common among AI tools. It also offered a token-priced Max Mode for top models.5 About a month later, the Pro plan itself moved to usage tied to API costs, which is closer to the token-style model the company had just said it was simplifying away from.

Features kept shipping

The product roadmap continued regardless. Version 0.50 added Background Agents that run tasks in parallel while developers work on other things, along with better context handling and faster inline edits.5 Versions 1.5 through 1.7, released in August and September 2025, added a Linear integration for launching agents from issue tickets, a /summarize command for long chat histories, agent autocomplete, and beta Hooks for monitoring and controlling agent behavior.5

The takeaway

The episode reads less as a one-off blunder and more as an early public collision between flat-rate consumer pricing and the variable cost of frontier AI models. Cursor's product momentum shows what users want. Agentic, always-on assistance is exactly the kind of usage that drives inference costs up. The company's real mistake was communication: it changed the terms for its most loyal customers without making the trade-off clear. Other AI coding tools face the same cost pressures, and the lesson for them is to be explicit about usage limits from the start, before users conclude they have been misled.