Cursor Composer 2.5 claims one‑tenth cost

Cursor’s latest pitch on Composer 2.5 rests on a familiar AI-industry formula: benchmark parity, lower price, and careful framing around where the comparison applies. TechTimes reported on May 20 that Cursor was positioning Composer 2.5 as matching Claude Opus 4.7 on coding benchmarks at roughly one-tenth the cost. Cursor’s own research index shows an “Introducing Composer 2.5” post dated May 18, 2026, and the company’s pricing docs place Composer 2.5 in its lower-cost “Auto + Composer” usage pool. The claim matters because Cursor is no longer selling only an editor that routes users to outside models. Cursor has been publishing research posts around its own Composer line since late 2025, and its March launch of Composer 2 described the model as a first-party coding system trained for long-horizon agent work. That gives the company an incentive to compare its in-house model directly with premium offerings from Anthropic and OpenAI. (techtimes.com) ### Where does the “one-tenth cost” line come from? TechTimes said the comparison centered on coding benchmarks and described the price gap as roughly 10-to-1 against Claude Opus 4.7. The outlet also said the comparison was selective and benchmark-dependent, rather than a blanket statement about all coding workloads. (cursor.com) Cursor’s own pricing pages support the broader point that Composer is positioned as a cheaper option inside the product. Cursor’s docs say the “Auto + Composer” pool is designed for “everyday agentic coding at a lower cost,” with Auto pricing listed at $1.25 per million input tokens, $6 for output, and $0.25 for cache read. The same docs separate that pool from the API-priced pool used for named frontier models from outside providers. (techtimes.com) ### Is Cursor actually saying Composer 2.5 beats Opus everywhere? Cursor’s public materials do not support that broader reading. The available record shows a benchmark-by-benchmark comparison strategy, not a claim of universal superiority across all software tasks. Cursor used the same style of framing when it introduced Composer 2 in March, saying it delivered improvements on the benchmarks it measured, including Terminal-Bench 2.0 and SWE-bench Multilingual. (cursor.com) TechTimes made the same limitation explicit in its May 20 report. The outlet said benchmark selection and selective metric comparisons were central to the story, and noted that cost comparisons can vary by task, dataset, configuration and workload. That means a token-price gap does not automatically translate into a 10-to-1 savings ratio on every real engineering job. (cursor.com) ### What does Cursor’s own benchmark history tell us? Cursor has been building the case for Composer through coding-specific evaluations for months. In March, the company said Composer 2 scored 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual, and said its scores were generated with specific harness choices and infrastructure settings. Cursor also disclosed that, for other models, it took the maximum of official leaderboard scores and scores recorded on its own infrastructure. (techtimes.com) That disclosure matters because benchmark outcomes in coding are sensitive to harness design, tool access, iteration count, and prompt setup. Cursor said Terminal-Bench 2.0 scores used the official Harbor evaluation framework for its own model, while Anthropic and OpenAI scores came from different agent harnesses. TechTimes’ caution about selective comparisons lines up with that kind of benchmark methodology detail. (cursor.com) ### Why compare specifically with Claude Opus 4.7? Claude Opus 4.7 is already a named premium option inside Cursor’s own product. Cursor announced Opus 4.7 availability on April 16 and described it in a forum post as “impressively autonomous and more creative in its reasoning,” while also offering a temporary 50% discount through April 22. That makes Opus 4.7 both a partner model inside Cursor and a direct pricing benchmark for users deciding which model to run. (cursor.com) Cursor’s current docs still separate Composer from those named API-priced models. The practical question for users is less whether one benchmark chart shows parity and more which model finishes a given repo task with the fewest steps, retries and output tokens. Cursor’s May 18 research index entry for “Introducing Composer 2.5” and its live pricing docs are the next places to watch for more detailed methodology or updated cost tables. (cursor.com) (forum.cursor.com)