Cursor, the company behind the AI code editor, released Composer 2 — its own frontier coding model, available in Cursor now. The company's one-line description: "Frontier-level coding with strong CursorBench results, higher token efficiency, and a faster default variant." Cursor says the model "delivers large improvements on all benchmarks we measure" and is "able to solve challenging tasks requiring hundreds of actions."
The number to write down is the price: $0.50 per million input tokens and $2.50 per million output — roughly a tenth of flagship pricing, for coding Cursor claims is frontier-level. Cursor calls the result "a new, optimal combination of intelligence and cost" and offers a "faster variant with the same intelligence."
One detail from Cursor's technical report is worth holding onto: Composer 2 was built by continued pretraining on Kimi K2.5, an open model, followed by large-scale reinforcement learning in realistic Cursor environments — the company's own product, turned into a training ground. That's the second story here, and we'll come back to it.
What a tenth of the price changes
Agentic coding is expensive in a way chat never was: the meter runs the whole time. An agent working a real task reads files, runs commands, hits errors, backs up, and tries again — Cursor's phrase is "hundreds of actions" — and every action burns tokens on both ends. At flagship rates, a long-running agent is a line item you notice. At $0.50/$2.50, a task that reads a million tokens of context and writes 200,000 back costs about a dollar.
That moves the build-vs-buy math on the unglamorous software a 2–50-person business actually runs on. The report someone assembles by hand every Friday. The CSV that gets massaged between the field-service app and the accounting system. The internal tool nobody built because it wasn't worth a developer's week. Capability was rarely the blocker on those projects — payback was. When the same afternoon of agent work costs a tenth of what it did, "too small to automate" gets redefined downward, and a pile of parked ideas quietly becomes viable.
The endurance claim matters as much as the price. "Hundreds of actions" is a statement about stamina — finishing a task, not answering a question. That's the axis the serious labs are competing on now, and it's the difference between AI that drafts and AI that does. It's also why how we scope engagements starts from tasks: a piece of work with a definition of done, small enough to verify, real enough to matter.
The second story is who shipped it. Cursor is not a research lab; it's a tool company that took a strong open model as a base and, per its technical report, trained it on the one thing only Cursor has — millions of real coding sessions inside its own product. That playbook keeps producing: Composer 1.5 argued that fast and cheap covers most everyday coding, Composer 2 argues that focused beats general on price, and Grok 4.5 pushes the same approach past software entirely. Expect more vertical labs — focused companies with deep workflow data fielding competitive models priced to win their niche. Where there are two, there will be ten, and the floor under this category will keep dropping.
What to do with this
- Re-price your parked automation ideas. Anything scoped and shelved because custom software cost too much deserves a fresh estimate. Run it through the three-question filter first — the price fell, but the need for a real, repeated task didn't.
- Scope tasks, not chat. A model built for "hundreds of actions" pays off on work with a definition of done: generate the report, reconcile the export, migrate the spreadsheet. Write the task down — inputs, output, what "done" looks like — before you point an agent at it.
- Treat every benchmark here as a vendor claim. All the performance numbers are Cursor's, measured on Cursor's own benchmark. The test that counts is one real task from your own backlog, measured on completion.
- Budget for retries, not perfection. At these prices the rational move is letting an agent attempt a task three times and keeping the best result. Cheap changes the workflow, not just the invoice.
If you want an honest read on which of your parked projects just crossed from "someday" to "this quarter," that's a 30-minute conversation.