Cursor shipped two releases two days apart: the Composer 2.5 model on May 18, 2026, and Cursor 3.5 on May 20. Together they lower the cost of a single agent call and raise the number of agents one developer can run at the same time.

Composer 2.5

Per Cursor’s changelog, Composer 2.5 is a step up from Composer 2 on sustained tasks and complex instructions. Cursor lists two token tiers: a Standard rate of $0.50 per million input tokens and $2.50 per million output, and a Fast tier at $3.00 per million input and $15.00 per million output. The Fast tier is the default, so the rate most developers pay without changing a setting is the higher one. The changelog does not publish a head-to-head benchmark against Opus 4.7 or GPT-5.5, so any parity claim circulating in coverage is not Cursor’s own.

Cursor 3.5 and multi-repo automations

Cursor 3.5, dated May 20, leads with multi-repo automations: agents can work across several codebases at once instead of one repository per session. Automations also move into the Agents Window, there is a no-repo mode for non-code tasks with a set of marketplace templates, and Cursor is running a time-limited discount on agent runs for newly created automations.

Where it fits

The two releases point the same direction. A cheaper Standard token rate makes each agent call feel inconsequential, and cross-repo automations multiply how many calls one developer sets in motion. Both are metered per run. The no-repo automations widen the surface further, pulling non-code chores into the same metered agent loop. Whether that combination is leverage or quiet overspend is taken up in parallel agents and the per-developer meter.