Cursor just beat EVERYONE.
Matthew Berman showed 6 AI tools in this video on . 5 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-22.
Tools shown in this video, with timestamps
Cursor's in-house coding model built on Moonshot's Kimi K2.5, scoring about 64% on CursorBench at roughly 55 cents per task, about 1.5 points behind Opus 4.7 Max at a twentieth of the cost. Priced at $0.50/M input and $2.50/M output, available only in Cursor, and trained with 25x more synthetic tasks than Composer 2.
Google's workhorse model released at Google I/O. On CursorBench it lands nearly 15 points behind the frontier and costs four times as much as Composer 2.5. The creator also discusses Sundar Pichai's reasons for Google's focus on Flash-class models.
Used as the frontier benchmark reference: top of CursorBench but the most expensive, at about $11 per task and $30 per million output tokens.
Nearly as good as Opus 4.7 on CursorBench but cheaper, about $4 per task at extra high. Praised for its price-to-capability ratio.
Open-source base model on which Composer 2 and 2.5 are built. Cursor drew criticism for under-disclosing this at first.
Model routing company that the creator invested in and calls the best on the planet. Mentioned as a way for enterprises to route workloads across models to control costs.
