Read this first
What this is. A reading of Cursor’s published pricing documentation, not a bill, not an invoice, and not financial advice. Every figure below comes from the vendor’s own pages, linked in the sources, all read on 2026-10-09.
What is verified twice. Prices quoted at provider list are taken from Cursor’s model table. The Claude Opus 5.5 figures were independently cross-checked against Anthropic’s own announcement.5 The GPT-5.6 figures could not be cross-checked from the machine that wrote this: OpenAI’s page returns a bot challenge to non-browser clients, so the only source read here was Cursor’s table.6 Treat the GPT-5.6 rows as Cursor’s numbers, not OpenAI’s.
What changes fast. Vendors reprice constantly. Cursor has already changed this fee once, and the plans it applies to have shifted. Assume the numbers are stale the moment you read them.
The short answer
Cursor does not charge a percentage markup on frontier model tokens. It passes the provider’s list price through and adds a flat fee on top of every third-party token: $0.25 per million tokens, charged on input, output and cached tokens alike.1
Because the fee is flat rather than proportional, its percentage effect is wildly uneven. On GPT-5.6 Sol output tokens it is a rounding error. On cache reads for a cheap model it is the entire bill.
The markup, in three parts
1. The Cursor Token Rate — $0.25 per million tokens
This is the actual markup. It is charged on input tokens, output tokens and cached tokens for any third-party model request, whether you pinned the model yourself or Auto and Cursor Router landed on it.1
- Teams and Enterprise only. The documentation attaches it to those plans specifically.1
- It covers search infrastructure, custom model execution and routing, and processing costs — Cursor’s stated rationale.1
- First-party models are exempt. Grok and Composer carry no token rate.1 The documentation is direct about the escape hatch: “How do I avoid the Cursor Token Rate? Use a first-party model, including Grok or Composer.”1
- BYOK does not exempt you. The rate applies to included usage, on-demand usage and BYOK usage alike, and stacks on top of whatever your API provider bills you.1
2. The old request-based plans: Max Mode at model rate + 20%
Legacy request-count plans carry a different structure: Max Mode bills at the model’s API rate plus 20%. New usage-based plans do not have it, and many of the newer models are marked as requiring Max Mode on those legacy plans.3 If you are on an old plan, this — not the token rate — is what you are paying.
3. Auto modes bill at the routed model’s price
There is no hidden multiplier for letting Auto choose. The request is billed at whatever model the router selected, plus the token rate if that model is third-party.1 The cost swings on which model the router picks, not on a surcharge.
What that does to specific models
Cursor’s listed price, the same price with the $0.25/M token rate added, and the percentage increase. Cache writes are listed for completeness; the fee is documented against input, output and cached tokens.
| Model | Input ($/M) | Output ($/M) | Cache read ($/M) | ||||||
|---|---|---|---|---|---|---|---|---|---|
| base | +fee | % | base | +fee | % | base | +fee | % | |
| GPT-5.6 Sol3 | $4.00 | $4.25 | +6.2% | $20.00 | $20.25 | +1.2% | $0.40 | $0.65 | +62.5% |
| GPT-5.6 Terra3 | $2.00 | $2.25 | +12.5% | $12.00 | $12.25 | +2.1% | $0.20 | $0.45 | +125.0% |
| GPT-5.6 Luna3 | $0.20 | $0.45 | +125.0% | $1.20 | $1.45 | +20.8% | $0.02 | $0.27 | +1250.0% |
| Claude Opus 5.55 | $4.00 | $4.25 | +6.2% | $20.00 | $20.25 | +1.2% | $0.20 | $0.45 | +125.0% |
| Claude Sonnet 5.53 | $2.00 | $2.25 | +12.5% | $10.00 | $10.25 | +2.5% | $0.10 | $0.35 | +250.0% |
| Claude Fable 5.13 | $10.00 | $10.25 | +2.5% | $50.00 | $50.25 | +0.5% | $0.25 | $0.50 | +100.0% |
Look at the last three columns. Cache reads are where agentic coding actually spends its money, and they are also where the flat fee does the most damage. On a cheap model’s cache read of $0.02/M, an extra $0.25/M is a 1,250% increase. The same fee on Sol’s $20/M output tokens is 1.25%.
A cache-heavy session, worked through
Agentic sessions are mostly cache reads: say 0.10M fresh input tokens, 0.90M cache reads, and 0.05M output tokens for one session.
| Model | Cost at Cursor’s list price | With the $0.25/M token rate | Increase |
|---|---|---|---|
| GPT-5.6 Sol | $1.760 | $2.022 | +15% |
| GPT-5.6 Terra | $0.980 | $1.243 | +27% |
| GPT-5.6 Luna | $0.098 | $0.361 | +268% |
| Claude Opus 5.5 | $1.580 | $1.843 | +17% |
| Claude Sonnet 5.5 | $0.790 | $1.052 | +33% |
| Claude Fable 5.1 | $3.725 | $3.988 | +7% |
On the expensive models the fee is noise — under 17%. On the cheap tier it nearly quadruples the session. This is the whole complaint in one table: a flat per-token fee is regressive, so it lands hardest on exactly the models teams adopt to save money.
The GPT-6.1 Sol problem
There is no Cursor markup on GPT-6.1 Sol, because Cursor does not appear to offer it at all. Cursor’s public model table stops at the GPT-5.6 family — Sol, Terra and Luna — and contains no GPT-6 entry of any kind.3
That makes the question moot for anyone on a subscription: you cannot select a model the picker does not list, so there is no markup to measure. The only conceivable route is BYOK, and Cursor’s own documentation restricts OpenAI keys to “standard, non-reasoning chat models” shown in its picker,4 which a GPT-6.1 Sol entry is not. Adding it would be a pass-through workaround rather than a supported path.
Bottom line, by plan
Individual plans (Pro, Pro Plus, Ultra)
Effectively no markup. Included usage is spent at provider list prices, and on-demand continues at the same rates. BYOK on an individual plan does not draw down included usage at all — your provider bills you and Cursor adds nothing.4
Teams and Enterprise
A flat $0.25 per million tokens on every third-party token — input, output and cached — including tokens bought with your own key. The measured effect runs from roughly +1% to +1,250% depending on model and cache mix.1
Legacy request-based plans
+20% on top of the model’s API rate whenever Max Mode is enabled, which many modern models require on those plans.3
Team seat pricing, for context
Standard seats are $40/user/month and Premium seats $120/user/month with five times the usage allowance. Usage does not transfer between team members and resets each billing cycle.2
One detail worth knowing if you are in India: Cursor’s documentation describes a Start plan, “our plan for developers in India”, which covers the first-party Cursor Models pool but does not include the third-party Other Models pool at all.3 That is a structural difference, not a discount.
What I could not verify
- OpenAI’s own list prices for the GPT-5.6 family. The page Cursor links to presents a bot challenge to non-browser clients from this machine, so those rows rest on Cursor’s table alone.6
- Whether the token rate applies to cache writes as well as reads. The documentation says “input tokens, output tokens, and cached tokens”; cache writes are most naturally read as input tokens, but the pages consulted do not say so explicitly.1
- Any real invoice. Everything here is list pricing against a hypothetical usage mix. An actual bill is the only way to settle what a given team pays.
Sources
Every figure above links here.
- S01 Cursor — Cursor Token Rate ($0.25/M, Teams and Enterprise, BYOK included)vendor documentation · primary · accessed 2026-10-09
- S02 Cursor — Team Pricing (seat types, included usage, token rate on third-party models)vendor documentation · accessed 2026-10-09
- S03 Cursor — Models & Pricing (full model table, list prices, usage pools, Start plan)vendor documentation · primary · accessed 2026-10-09
- S04 Cursor — Bring your own API key (BYOK rules per plan, OpenAI key restrictions)vendor documentation · accessed 2026-10-09
- S05 Anthropic — Introducing Claude Opus 5.5 ($4/$20 per M, $0.20 cache reads)vendor announcement · independent cross-check · accessed 2026-10-09
- S06 OpenAI — Previewing GPT-5.6 Sol (linked from Cursor’s model table; page bot-gated from this machine)linked source · not read directly · accessed 2026-10-09