I held a legacy Cursor plan for a year. Its own docs say it expires at my next renewal

I have been paying Cursor twenty dollars a month since before twenty dollars a month meant what it means now.

For a while that was worth something. "I'm on the legacy plan" was a complete answer to why I had never gone looking at alternatives — there was something in my seat a new customer could not buy, so the question stayed closed and I renewed on autopilot.

Last week I went and read Cursor's own documentation end to end, because I had written a sentence about their pricing in another article and wanted to check it before it aged badly. On the page describing the plan I thought I was on, under a heading called Legacy customers, I found this:

If you're on a legacy request-based plan, you can continue using it until your next renewal. At renewal, you'll be migrated to the usage-based pricing model.

Not a deprecation notice with a date on it. Not an email. A sentence in the docs, and a trapdoor under the renew button.

So here is the whole arc, sourced end to end from Cursor's own blog and docs, and an honest audit of what a grandfathered seat is worth in August 2026. The short version is the part I did not expect: less than a new one.

What I actually bought

Cursor Pro, before June 2025, was a request plan. Michael Truell's own post spells out the terms: "There was a limit of 500 requests per month, with Sonnet models costing two requests."

That is the entire spec, and the important thing about it is the unit. A request was a request. It did not matter how much of your codebase rode along behind it, how long the session had been running, or how many files the model read before it answered. You bought turns, not tokens.

For the way I work that was an unreasonably good deal, and I want to be precise about why rather than nostalgic. Under a request plan, thinking harder is free at the margin. The rational move is to make every turn as heavy as you can — attach the whole module, paste the failing test, let it read everything — because a fat request and a thin one cost the same one credit. Context was the thing you were not paying for.

That is a strange economy, and I understood at the time that it could not survive an agent that reads forty files before it answers. But it was the economy I bought into, and 500 requests was the number on the tin.

June 16, 2025 — the unit changes

Cursor announced Ultra, a $200/month tier, and changed Pro underneath it:

By default, the Pro plan will move from request limits to compute limits; all users will get at least $20 of model inference at API prices per month. Concurrent with this change, we're rolling out unlimited access to the "Auto" model and lifting all limits on tool calls. Existing users can choose to stay with the "500 request limit" method if they prefer.

That last sentence is the legacy plan. It is one clause, at the end of a paragraph, in a post whose headline was about a new $200 tier. If you were an existing user in June 2025, that clause is the entire documentation of the choice you were being offered.

The new default is worth reading closely, because the two halves do different jobs:

The floor is unmetered — Tab completions, and anything Auto routes for you. Above the floor you buy tokens at the provider's list price, with $20 of them included. Which is to say the subscription price and the credit became the same number.

July 4, 2025 — the apology

Two and a half weeks later, Truell published Clarifying our pricing. It opens:

Our recent pricing changes for individual plans were not communicated clearly, and we take full responsibility. We work hard to build tools you can trust, and these changes hurt that trust.

And names the specific sentence that did the damage:

We were not clear that "unlimited usage" was only for Auto and not all other models, which have at least $20 of included usage.

With money attached and a dated window: refunds for unexpected usage between June 16 and July 4, on request to pro-pricing@cursor.com.

I want to give credit where the rest of this article will not. That is a good apology. It identifies the misleading clause instead of blaming the reader for misreading it, it publishes a timeline including the quiet June 30 edit to the original post, and it comes with a refund address rather than a goodwill gesture.

It also contains the one number nobody quotes:

Based on median token usage, the Pro plan currently covers about 225 Sonnet 4 requests, 550 Gemini requests, or 650 GPT 4.1 requests.

Hold that against the old plan. Five hundred requests, at two per Sonnet call, is 250 Sonnet requests. The new plan covered about 225. At the median, the change cost a Pro user roughly ten per cent — genuinely modest, and Cursor was right to say most people would not notice.

But the median is doing enormous work in that sentence, and the same post tells you why: "the hardest requests cost an order of magnitude more than simple ones." Under a request plan, that variance was Cursor's to absorb. Under a token plan it is yours. The median user lost ten per cent. The user whose requests are the order-of-magnitude ones lost a great deal more — and that user is precisely the one for whom the old plan was worth defending.

Which brings us to what everyone was told they still got for free.

August 12, 2025 — the concession is withdrawn

Unlimited Auto is what made the June plan survivable. It was the shock absorber: the place you did the work that did not need a frontier model, so the $20 pool could be spent on the turns that did. And in the July 4 post it sits at the top of the list of what the new Pro gives you.

Five and a half weeks after the apology, Cursor announced that it was going away:

At your next billing renewal after September 15, Auto will contribute to your included monthly usage at competitive token rates. From December 2023 to June 2025, Auto was priced at the same cost as other premium models. Since June 2025, Auto has been unlimited for individuals.

Read the dates in that quote in order. Auto was metered until June 2025. Auto was unlimited from June 2025. Auto is metered again from your next renewal after September 15, 2025.

For a monthly subscriber, unlimited Auto existed for about three months. It arrived as the sweetener on a change people were angry about, it was held up on July 4 as the reason the change was fair, and it was withdrawn on August 12. Annual subscribers kept it longer — the post is explicit that the change lands at your renewal, so "if you bought a yearly subscription at June 2025, these changes would only take effect at June 2026" — which is why you will find people online swearing Auto was free well into 2026. They are not wrong. They were just on a different clock.

None of this was hidden. All three posts are still up. But it is worth noticing that the thing offered as redress in July was on a schedule to be removed in August, and that no post ever connects the two.

Where that leaves a legacy seat, August 2026

I could keep narrating, but the current docs do it better. Here is the model table on the legacy pricing page — the page for the plan I have been paying for — with the column that matters:

Every current Claude, every current GPT, every Grok: no request price. A dash. My 500 requests buy Gemini and Composer, and nothing else.

The frontier models are still reachable, and the same page says how:

Max Mode is available at the model's API rate plus 20%. Max Mode enables larger context windows, subagents, image generation, and access to the latest frontier models on request-based plans.

So Max Mode — the one feature the current pricing docs confirm is "available only on legacy request-based plans" — is not a perk. It is the toll booth. It is the only road from a legacy seat to Claude Opus 5, and it charges twenty per cent over the API rate to use it.

Now put the two $20 plans side by side.

The grandfathered plan is strictly worse. Same price, narrower included usage, and a twenty per cent surcharge on exactly the models anyone is actually paying Cursor to reach in 2026. Whatever "legacy" meant in June 2025, it now means paying more for less, and it terminates at my next renewal regardless.

I do not think that is cynical on Cursor's part, and I am not going to pretend it is a scandal. Max Mode was designed inside a request economy, where a bigger context window genuinely was a distinct product that needed its own price. Carrying it into a token economy required attaching a token rate to it, and API + 20% is a defensible one. Every step here has a reasonable explanation. It is just that the sum of the reasonable steps is a plan nobody would choose.

The arithmetic that actually decides it

Price the alternative honestly, because this is where a good product usually wins.

Claude Opus 5 through Cursor is billed at $5 per million input tokens and $25 per million output — Anthropic's list price, unmarked up, on a current Pro plan. The $0.25-per-million Cursor Token Rate that people quote applies, per the docs, on Teams and Enterprise plans; on Pro you are simply buying tokens at cost.

That sounds like a reason to stay, and for a year it was. But it means the subscription is not buying me a price any more. It is buying me a product, and the product now has to win on its own merits — because there is no longer a discount underneath it to make the argument for me.

There is also no structural edge left for a reseller to hold. The one a middleman could plausibly have had is long context: buy it in bulk, smooth it, resell it. That arbitrage does not exist on Claude. Anthropic's pricing page is explicit that models from Claude 4.6 on "include the full 1M token context window at standard pricing," and that "a 900k-token request is billed at the same per-token rate as a 9k-token request." No cliff, no tier, nothing to smooth. (OpenAI's recent models do price long-context input at a premium, so this is a Claude-specific fact rather than a law of nature.)

And Cursor's own docs tell you what the $20 actually covers now:

Daily Agent users: Typically $60–$100/mo total usage

Power users (multiple agents/automation): Often $200+/mo total usage

That is the company's own estimate, on its own pricing page. If you use an agent daily, the $20 plan is not the price of the product. It is the minimum, and the bill is three to five times larger. Which is fine — that is what the tokens cost, and Cursor is passing them through — but it settles what kind of decision this is. I am not choosing between a $20 subscription and paying API rates. I am choosing where to route a $60–$100 monthly token spend, and the subscription is a $20 access fee on top of it.

The Hong Kong footnote

Two things make this land differently here.

The first is that our own Cursor in Hong Kong guide, written in March, tells readers that Auto is the right place to start. That was good advice when it was written, and part of why it was good advice is that Auto was where the metering was not. It needs an asterisk now, and this article is the asterisk: Auto is metered, all three modes bill at the list price of whatever model they route to, and "just use Auto" is no longer a way to avoid thinking about the bill. The guide's other point stands and has aged the other way — Cursor's Claude access has been reported blocked for some HK and China-region users, which is exactly the kind of dependency worth pricing before you renew.

The second is data residency, which is a live question for anyone here handling client code. Cursor's pricing page notes that opting into regional data residency "incurs a 10% uplift on Model pricing," and Anthropic's page independently confirms a 1.1× multiplier on all token categories when you pin an inference geography. Both layers charge the same ten per cent for the same thing. Worth budgeting for rather than discovering.

What has moved in the other direction is the exit. Two years ago, leaving Cursor meant giving up agentic coding. It does not any more: an Anthropic-compatible endpoint is a commodity. Qwen Cloud's Token Plan hands you a base URL and a key, DeepSeek does the same, and anything that speaks the protocol hangs off it. I spent a full week working that way when my Claude Code quota died, and wrote down what it cost to the token. The relevant finding for this article is not the price. It is that the switch was a symlink and a restart.

So am I cancelling?

Not this month, and I would rather be straight about why than write a tidier ending.

I am still paying because Tab is genuinely good, because the editor is where my hands already are, and because one bill is easier than four. Those are real reasons. They are also, every one of them, reasons about the product — not one of them is a reason about the plan.

That is the actual change, and it is smaller and more final than a price rise. The plan used to argue for itself. "I'm on the legacy tier" closed the question, because there was something in the seat I could not get by leaving. Now Auto is metered like everyone else's, tokens cost what they cost, the one exclusive I have left charges me twenty per cent over the API rate to use it, and Cursor's own documentation says the whole arrangement ends at my next renewal anyway.

So the subscription has to win the argument every month now, on the merits, against a symlink. It might keep winning — that is a real possibility and I am not predicting otherwise. But nothing is holding it in place any more, and the gap between those two states is the gap between a customer and a habit.

If you are an old-timer too, the check takes a minute. Open the model selector: if Max Mode is there, you are on a legacy request-based plan, and the docs above are describing your seat. Then look up your renewal date, because that is the day it stops.

Sources are all first-party. Cursor's June 16, 2025, July 4, 2025 and August 12, 2025 posts for the timeline and every quotation attributed to Michael Truell; Cursor's Models & Pricing and Request-Based Pricing (Legacy) docs for current rates, the legacy model table, Max Mode and the migration-at-renewal language; Anthropic's pricing page for model rates, long-context pricing and the data-residency multiplier. All read on 2026-08-31; prices and terms move, so check them against the live pages before acting on any of it. My own seat: a personal Pro subscription bought before the June 2025 change.

Get notified when we publish new articles and episodes. No spam, just signal.