The .tsx is a Cursor IDE canvas (TypeScript + React). A browser cannot render it — save the file and open it in Cursor to see the live preview.
Every Claude and GPT row. Burns included third-party credit at that model’s API price. Live docs say the pool is included on Pro and up; they do not publish a dollar figure. Hidden rows still bill if you turn them on.
The licence is a monthly floor. Token estimates in the next section are extra once included usage is exhausted (on-demand, billed in arrears). Teams add $0.25/M on Claude and GPT.
Choose a job and a size. The chart is the estimated USD for that one job at published Cursor rates. Heavy software uses Cursor’s own CursorBench 4.0 cost-per-task where they published one.
CursorBench 4.0 grades ambiguous, multi-file work from real Cursor sessions. A 10-point gap is meaningful; 0.3 points is noise.
EQ-Bench Creative Writing v3 ranks short-form stories by pairwise Elo. Treat Elo as a writing taste test, not a Cursor coding rank.
Jane Street’s head of quant research said Fable 5.1 is state of the art on their trading-intuition tests. That is a customer quote, not a public leaderboard. No model here sees live bids or your broker. For trading work in Cursor: Luna / Haiku / Grok to ingest; Opus or Fable to argue the thesis; you to take the risk.
Source: cursor.com/docs/models-and-pricing. Sol promotional pricing through 21 Nov 2026. Grok Fast not shown (2×).
Two pictures. The first is 20 standard features + 8 research memos + 4 short stories. The second is 20 hard CursorBench 4.0 tasks only — closer to a power-user coding month.
Formula from the Standard mixes. Cache reads included on the coding slice.
20 × published CursorBench 4.0 $/task. Haiku omitted (not measured).
Every Grok, Anthropic, and OpenAI row on Cursor’s public model table as of this snapshot. Hidden models still work if you enable them; they still bill. Composer 2.5 is a Cursor model, not Grok/Claude/GPT (CursorBench 4.0: 27.7% at $0.68/task).