Cursor Composer 2.5

Cursor has spent the last three years turning a VS Code fork into something that no longer feels like one, and Composer 2.5 is the clearest sign of that shift. Launched in May 2026, it's the third generation of Cursor's own agentic coding model, and rather than chase the top of the intelligence leaderboards it goes after a different prize: doing the vast majority of real engineering work fast and cheap enough that you stop thinking about the meter.

The approach is unusual and worth understanding before you judge it. Composer 2.5 isn't trained from scratch — it's built on Moonshot AI's open-source Kimi K2.5 checkpoint, with Cursor reportedly pouring 85% of the model's total compute budget into its own post-training. That's a pragmatic bet: let someone else pay for the base model, then spend everything on reinforcement learning against the specific loop that matters inside an IDE — read files, edit across a repo, run terminal commands, run the tests, fix what broke, repeat.

It's also not the only model on offer, which matters a great deal. Cursor still fronts Claude Sonnet 4.7 and Opus 4.7, GPT-5.5 and GPT-4.1, Gemini 2.5 Pro and Grok 4, plus the Sonic Tab model for low-latency completions. Composer 2.5 is the workhorse you reach for by default and escalate away from when a problem genuinely warrants it. Available across desktop, web, CLI, mobile and SDK, and reachable from the JetBrains plugin and Background Agent, it's aimed squarely at professional developers who live in a large codebase and want an agent running long-horizon tasks without a finance conversation attached.

The benchmark picture is genuinely strong, and more importantly it's strong in the right places. A score of 62 on the Artificial Analysis Coding Agent Index puts Composer 2.5 third overall, behind Claude Opus 4.7 in Claude Code (66) and GPT-5.5 at xhigh reasoning in Codex (65). Third place is respectable on its own. Third place at $0.07 per task, against $4.10 and $4.82 for the two models above it, is the actual headline — roughly 10x to 60x cheaper for a four-point gap. For most day-to-day work, that trade is not close.

The generational jump is real rather than cosmetic. The +35 point leap on SWE-Bench-Pro-Hard-AA, from 12% to 47%, is the kind of movement that changes what you can actually delegate; Composer 2 simply could not be trusted with hard multi-file problems, and 47% means you can now hand one over and expect a coin-flip-or-better result. The gains elsewhere are much more modest — two points on Terminal-Bench v2 and three on SWE-Atlas-QnA — which suggests the post-training was tightly targeted at difficult repository-scale editing rather than broadly improving the model. Fair enough, but worth knowing: if your bottleneck is terminal work or codebase Q&A, this upgrade will feel far less dramatic than the marketing implies.

Fast mode is where I'd push back hardest. Averaging 6.7 minutes of wall time per task makes it the third-fastest agent on the Index, but you pay roughly 6x per task for about 30% less waiting. The token pricing tells the same story: $3.00/$15.00 per million versus $0.50/$2.50 for standard. That's a poor ratio for anything you aren't actively watching, and given Composer 2.5's entire pitch is long-running background work, Fast mode undercuts its own value proposition. Use it for interactive iteration, leave it off for anything you've queued and walked away from.

Two caveats temper my enthusiasm. First, the 200,000-token context window is adequate but no longer generous, and it's a real constraint for an agent whose main job is reasoning across many files at once — expect to lean on project rules and scoped tasks rather than throwing a monorepo at it. Second, the metered credit model across Free, Pro ($20), Pro+ ($60) and Ultra ($200) makes costs genuinely hard to forecast. Cheap per task is not the same as predictable per month, especially if your team habitually escalates to frontier models. The claim that Composer 2.5 nearly matches Opus 4.7 on SWE-Bench Multilingual (79.8% vs 80.5%) comes from Cursor itself and I'd treat it with more caution than the independently measured figures.

None of that changes the verdict much. Composer 2.5 is the most economically sensible default coding agent I've assessed, wrapped in an editor that gives you the expensive alternatives one dropdown away when you need them. At $20/month for Pro, this is an easy recommendation for working developers; the reservations are about pricing transparency and context headroom, not capability. It earns 4.5 out of 5.