Cursor is losing roughly $32 a month on every Pro subscriber. That’s not a leak. That’s what the public math says if you read their July 2025 apology carefully and put a calculator next to a Sonnet 4.6 rate card.

The $20 plan was always a tripwire. Nobody seems to be talking about what tripped.

I spent an afternoon with a spreadsheet, a stack of Cursor blog posts, and the Anthropic pricing page open in three tabs. The numbers don’t argue with each other. They just sit there and stare.

What the July apology actually admitted

On July 4, 2025, Cursor CEO Michael Truell posted a short note on the company blog. The headline was “Clarifying our pricing.” The substance was an apology for switching the Pro plan from 500 fast requests to a $20 credit pool on June 16.

Read between the lines. The reason they had to change Pro to credits is that the prior plan was uncapped against Sonnet 4 — and people noticed they could run it 18 hours a day.

Two important things in that post if you squint. One: the $20 plan was always going to include “unlimited” usage on Auto — Cursor’s internal router that picks the cheapest model that can still finish the task. Two: every other model, including the one your power users actually want (Sonnet 4.6, now 4.7), comes out of a $20 credit pool that gets eaten in about 225 prompts.

Translation. The product they market is unlimited Sonnet. The product they ship is unlimited GPT-4o-mini-tier Auto plus a coupon for Sonnet.

The math nobody at Cursor wants you to do

Let’s run the numbers as they exist in public.

Anthropic’s posted rate for Sonnet 4.6 is $3 per million input tokens and $15 per million output tokens. A typical Cursor “agent turn” sends roughly 30K tokens of context (codebase, repo map, conversation history) and gets back about 4K of code. That’s $0.09 in plus $0.06 out. Eleven cents per turn.

The median Cursor Pro user, according to multiple third-party teardowns of usage logs floating around dev twitter, runs 200 to 600 turns a day across model-routed and Sonnet-direct calls. Take the conservative end. Take a quarter of those as Sonnet-direct (because Auto handles the rest). That’s 50 Sonnet calls a day, $5.50 in raw COGS, or about $165 a month.

Power users — and Cursor markets to power users — easily double that. The user pays $20.

Even using the Cursor-published “median” of 225 Sonnet calls a month at 11 cents each, you get $24.75 in pure inference cost. That’s already underwater before you pay rent, salaries, support, or the Stripe fee.

Now layer in everything Cursor doesn’t show you. The retrieval calls (every keystroke, embedded, re-ranked). The model swaps when Auto bounces to Opus. The retries on bad diffs. Brutal.

Why Anthropic-only routing is the real story

Open Cursor on a fresh install. Look at the model picker. Sonnet 4.6 (or 4.7, depending on when you read this) is the default. Auto routes most prompts through Anthropic. Even the cheap fallbacks lean Claude-heavy because Cursor’s own evals say Anthropic outperforms on agentic code tasks.

This is a problem.

When 70 to 90 percent of your prompts route through one vendor, your gross margin is whatever that vendor decides it should be. If Anthropic raises Sonnet by 25% next quarter — which they have done twice in the last 18 months, including a quiet “tier rebalance” in March — Cursor either eats it or shoves the bill onto the user. They cannot diversify away.

The July 2025 apology framed the pricing change as a “communication issue.” It wasn’t. It was the first visible sign that the unit economics had stopped working. The credit pool was a margin defense disguised as a UX tweak.

The timeline nobody is connecting

1. Jun 16, 2025Cursor silently swaps Pro from “500 fast requests” to a $20 credit pool. Twitter erupts.
2. Jul 4, 2025Truell posts the apology. Refunds offered for Jun 16–Jul 4 surprise charges. The credit model stays.
3. Aug 2025Teams pricing gets a quiet update. Org admins now see per-seat usage caps for the first time.
4. Q1 2026Anthropic pushes Sonnet 4.7. Cursor makes it default. Inference costs go up across the fleet.
5. May 2026Annualized revenue floats around $1B against COGS of ~$1.15B. Reported widely. Confirmed by nobody.

None of this is a leak. Every one of these data points is in a public blog post, a press release, or a quote on the record. The story is what happens when you put them next to each other.

What I’d bet a price hike looks like

Cursor cannot raise the headline number on $20 Pro. Doing so would acknowledge the apology was empty and trigger a second churn wave (the first one, in July 2025, was bad enough that the company published a follow-up explaining the refund policy).

So they will do what every SaaS company in margin pain does. They will quietly shrink the included usage and call it “rebalancing.” They will introduce a new tier above Pro — call it “Pro Max” at $40 — that gets you Sonnet on the front-end model picker without the credit accounting. They will route more aggressively to Auto and pretend that’s a quality improvement.

If they’re feeling brave, they ship a “bring your own API key” mode and let the user eat the inference bill directly. Cline already does this. Aider always did. Cursor has resisted because BYOK kills the data moat and the retention story.

My best guess on timing: a price-hike-shaped event before Q4 2026 earnings. Either a tier above Pro, or a tighter credit cap on Pro itself, sold as a feature. If neither shows up, Anthropic will eventually do them a favor and raise rates again — which means the same thing with a different villain.

The blind spot I haven’t solved for

I’m assuming Cursor isn’t getting a meaningful discount from Anthropic. Probably wrong. Cursor is rumored to be one of Anthropic’s three biggest API customers — somewhere north of $400M in committed spend. There is almost certainly a volume tier nobody outside both companies sees, and it could be 30 to 50 percent off rack rate.

That doesn’t fix the problem. It just changes the slope. At 40% off, the $32-per-user burn becomes $12-per-user burn. Still negative. Still unscalable. Still a hostage situation where the hostage taker prints the discount sheet.

(I also haven’t tested whether retrieval-side caching gets them another 10 to 15 percent. It probably does. It doesn’t change the structural argument.)

The recast belief

Cursor is not a SaaS company. Cursor is the world’s best-funded reseller of someone else’s API, wrapped in an excellent UI. The UI is the moat right up until the moment Anthropic ships a competing IDE — and Anthropic has been hiring frontend engineers all spring.

If you’re a founder building a wrapper on Sonnet right now, this is the math you need to internalize. Your COGS is a politburo decision. Plan your runway like it.