When Cursor's $20 Plan Emptied in an Afternoon

The afternoon the wallet ran dry

Picture opening Cursor on a Tuesday, firing off a handful of prompts to a coding agent, and watching a brand-new month's allowance evaporate before lunch. That's roughly what a wave of Cursor Pro users described in late June 2025 — and a few weeks later the company's own CEO sat down to write an apology. I find these stories useful precisely because nobody did anything wrong. No runaway loop, no forgotten background job. Just developers using a tool the way they always had, after the meter underneath it quietly changed shape.

What changed on June 16

For most of its life, Cursor's Pro plan sold you requests. As the company later explained, there was a limit of 500 requests per month, with Sonnet models costing two requests each — so in practice you got on the order of 225 heavy Sonnet turns and then slower, unlimited fallback. Predictable. Boring. Easy to budget. On June 16, 2025, that flipped. The new Pro plan gave you, in Cursor's own words, "$20 of frontier model usage per month at API pricing" plus unlimited usage of Auto mode. The catch hiding inside that sentence: only Auto mode stayed unlimited. Every time you hand-picked a frontier model — exactly what power users do — you were now spending real API dollars out of a $20 wallet. TechCrunch summarised the swap cleanly: instead of 500 fast responses on advanced models and then unlimited slower ones, subscribers would now get $20 worth of usage per month, billed at API rates. Same $20 price tag on the box. Completely different thing inside.

Why $20 didn't go nearly as far

Here's the mechanism, because the number alone doesn't explain the shock. Under request-based pricing, a "request" was a flat unit — one Sonnet turn cost two of your 500, full stop, no matter how much it actually computed under the hood. Under API pricing, you pay for tokens: every token you send in and every token the model writes back. And an agentic coding turn is enormous on the input side. The agent re-sends the system prompt, the context window full of your open files, prior turns, tool definitions and tool output — on every step. Cursor itself conceded the point: "new models can spend more tokens per request on longer-horizon tasks", and the hardest requests cost an order of magnitude more than the simple ones. So the $20 stretched very differently depending on which model you reached for. Independent breakdowns of the new plan put it at roughly 225 Claude 3.5 Sonnet requests, 500 GPT-4o requests, or 550 Gemini requests for your $20 — and far fewer with the newest, hungriest Claude models. The exact thing that made Cursor great for coding, Anthropic's Claude models, was the exact thing that drained the wallet fastest. That's the trap. Nobody budgets in tokens. They budget in "I'll do a few hours of coding." When the meter underneath is token-based and your agent re-reads the whole repo each turn, "a few hours" and "$20" stop being the same sentence.

What it actually cost people

The receipts were not subtle. Users reported running out after just a few prompts when using Anthropic's new Claude models, and documented cases of a subscription depleting in a single day of normal usage circulated widely. To ground the token economics: at the time, Anthropic's flagship Opus 4 / 4.1 listed at $15 per million input tokens and $75 per million output tokens, with Sonnet at $3 / $15. At those rates, a $20 wallet is barely a few million output tokens of Opus — and an agentic session re-stuffing a fat context window each turn chews through input tokens at a frightening clip. The token-cost math for AI coding agents is brutal once input dominates the bill, which it almost always does. (For the underlying per-model rates, our LLM API token pricing breakdown is the reference.) The backlash got loud enough that on July 4, 2025, Anysphere CEO Michael Truell published a public mea culpa: "We recognize that we didn't handle this pricing rollout well and we're sorry. Our communication was not clear enough and came as a surprise to many of you." The company offered refunds for unexpected usage between June 16 and July 4, pointing affected users at a dedicated pricing-support address.

What token discipline would have changed (an estimate)

Let me be honest about what is and isn't fixable here. You cannot opt out of someone else's pricing change. What you can change is how many tokens your agent burns to do the same work — and under the new token-based meter, that's the whole game. The dominant cost in agentic coding is input: the same files, history and tool output re-sent every turn. That's where the lever is. Concretely:

• Don't feed the model whole files. Reading a 400-line file to confirm one function signature is pure waste. Sending a skeleton or the matched symbol instead can cut a read by 80–95%.
• Cache the stable prefix. With prompt caching, the unchanged system prompt and project context cost a fraction on every subsequent turn instead of full price each time.
• Trim verbose tool output before it ever lands in context.

These are exactly the levers in our guide to reducing AI coding agent token usage. As a labelled estimate, not a measured result: if input tokens are 80–90% of an agentic bill — which they typically are — and disciplined retrieval plus caching trims input by a conservative 40–60%, then a session that drained the $20 wallet in an afternoon would have lasted on the order of two to three times longer for the same work. That doesn't undo the pricing change. It does mean the meter ticks far slower against the same $20.

The lesson: the meter changed, so change the burn

The Cursor episode isn't really a story about a company that priced badly and apologised — they did, and they did. It's a story about what happens to everyone the day a flat-rate tool becomes a token-metered one. The work didn't get more expensive. The accounting got honest. And honest token accounting punishes the way most of us let agents work — re-reading everything, every turn, in full. You can't control when Anthropic, OpenAI or a tool vendor reprices the meter. You can control how many tokens cross it. Tokenade sits between your coding agent and the model and does the unglamorous part automatically: skeletons instead of whole files, compacted tool output, cache-friendly prefixes — so the same work sends a fraction of the tokens. When the next pricing change lands, a leaner burn is the only hedge that's actually in your hands. Start free → — and the next time the meter shifts under you, you'll feel it a lot less.