TEARDOWN Nº 2 / WHAT A CURSOR SESSION COSTS

What a Cursor session costs, by model

PUBLISHED 2026-08-13 · BY Llorenç Ballester — Chief Burn Officer

Cursor's own documentation says on-demand usage "continues at the same API rates" once your monthly pool runs out. Take that at face value and it turns Cursor Pro's $20 credit into exactly what it sounds like: a raw API budget. We already price a coding-agent session across every model on our board — so here is what $20 of Cursor buys you, by model, and why the answer swings 123x.

The $20 pool is a real budget, not a request count

Cursor Pro is $20/month, and per Cursor's own pricing docs (captured 2026-08-13), that plan includes “$20” of “Other Models usage included,” and once it's gone, on-demand usage “continue[s] at the same API rates with pay-as-you-go billing.” That is a specific, checkable claim: Cursor is telling you the pool is priced in the same dollars the model provider charges, with no markup on the Individual plan. (Teams and Enterprise add a $0.25 per-million-token surcharge on third-party models — a different lane, not covered here.)

Which means the number that actually decides how far $20 goes isn't a Cursor number at all. It's the model's own API price — which is exactly what our board tracks daily.

One session, 17 prices

We don't have Cursor's own token telemetry — nobody outside Cursor does. What we have is our own published canonical coding-agent workload: a 15,000-token aggregated input, 1,500-token output, 40% of the input served from cache — the same shape this site already prices at /workloads, unchanged for this article. Run it against every model and this is what one session costs:

Prices from the board, last confirmed 2026-08-12. “Monthly at default volume” uses this workload's own published assumption of 880 sessions/month (40/day × 22 workdays) — swap in your own volume at /workloads. Estimates at list prices, not invoices, and not Cursor's own accounting.

11 of 17 models blow past the pool at this volume

Nobody runs exactly 40 agent sessions a day on one model forever — this is a fixed yardstick, not a prediction of your bill. But held constant, it makes the spread concrete: 11 of the 17 models we track would burn through a $20 pool before the month is out at this shape and volume, led by Claude Fable 5 at 9.2x — call it 3 days of the month before on-demand billing takes over. The other end of the board, DeepSeek V4 Flash, uses 7% of the same pool for the same monthly volume — the $20 outlasts the month with room to spare.

Manually pinning the flagship model for every step is not a bug in Cursor's pricing — it's the price of that model, passed through honestly. What it means for you is that “which model is selected” is the single biggest lever on how far a fixed monthly credit goes, well before prompt length or session count start to matter.

The caching wrinkle, again

Our first teardown found that Anthropic models won't cache a prefix under a published minimum — 512 on Claude Fable 5, 512 on Claude Opus 5, 1,024 on Claude Opus 4.8, 1,024 on Claude Sonnet 5, 4,096 on Claude Haiku 4.5. It applies here too, with a sharper edge: caching references “the entire prompt — tools, system, and messages — up to and including the block designated with cache_control,” and a request can carry several such breakpoints, each needing to independently clear the minimum at its own cumulative length. An agent tool that caches tool definitions, system prompt and context as separate breakpoints can have some of them land above the line and others below it — nothing in the response tells you which, unless you read the cache token counts yourself.

On Claude Haiku 4.5, our own workload's 40% cache-hit assumption is a best case. If none of it clears the model's 4,096-token minimum, the same session goes from $0.0209 to $0.0262 — 26% more, silently, for a model that is otherwise the cheapest way to run a real coding agent on Anthropic's board.

What to actually do with this

1. Know which lane you're in. Auto mode and manual model selection are priced differently in Cursor; only manual selection draws from the $20 pool at the model's own rate, which is the only lane we can price from public numbers.
2. Price your own shape, not ours. Our 15,000-in/1,500-out session is a yardstick, not your codebase. Swap in your real prompt size and session count at /workloads.
3. Route deliberately, not by default. If the flagship model is 123x the cheapest one for the same shape of work, picking it for every step is a choice with a price tag — see model routing for when the expensive model is actually worth it.
4. Watch the cache counters, not just the bill. On Anthropic models, confirm cached tokens are actually being read — see the cache-minimum teardown for exactly how.

Method and limits

Every session price is computed at build time from the pricing board and the coding-agent workload — nothing here is Cursor's own accounting, and we have no visibility into Cursor's actual per-request token usage. The one number in this article that is not ours — the $20 pool and its at-cost overage policy — is Cursor's own stated policy, cited above and dated, and it is the one figure on this page that our own daily watcher does not track: if Cursor changes it, this article goes stale until we notice and update it by hand. Everything downstream of that one fact — the session prices, the spread, the pool math — is recomputed from the board on every deploy. Full rules at /methodology.

Frequently asked questions