Cursor just shipped Composer 2.5 - a coding model that matches frontier models at a fraction...
Cursor released Composer 2.5 this week, a coding model that the company says matches frontier general purpose models on software engineering benchmarks while costing a small fraction per task. For anyone tracking the economics of AI coding tools, this is one of the clearer signals yet that specialized models are catching up to the giants on the work that actually pays the bills.
The training approach is the interesting part. Cursor trained Composer 2.5 on twenty five times more synthetic tasks than its predecessor and added a reinforcement learning feedback loop where the model learns from its own mistakes during a task rather than only at the end. The analogy is a student who gets detailed notes on exactly where their reasoning went wrong, rather than only a final grade weeks later. This kind of dense, in process feedback has been one of the quieter trends in model training over the last year, and it tends to produce models that are much better at multi step work like editing a codebase, running tests, and responding to failures.
The numbers back it up. Composer 2.5 scores 79.8 percent on SWE-Bench Multilingual, up from 73.7 percent in the previous version. That puts it shoulder to shoulder with Claude Opus 4.7 and GPT-5.5 on coding benchmarks. The cost story is where it gets more striking. Cursor says Composer 2.5 runs at under one dollar per task, while competing frontier models can charge up to eleven dollars for the same work. For teams running thousands of agent tasks a day, that is the difference between a tool you experiment with and one you actually deploy at scale.
There is also a longer term bet underneath this. Cursor is partnering with SpaceXAI to train a much larger model from scratch on the Colossus 2 cluster, using roughly ten times more compute than what produced Composer 2.5. So the strategy is not to stay small forever, but to ride the efficiency curve now while scaling up in parallel. The gap between a frontier general model and a purpose built specialized one is closing fast on coding tasks specifically, and the cost curve is bending at least as fast as the capability curve.
What to watch next is whether this pattern holds in other narrow domains, legal drafting, medical reasoning, scientific literature work, where a focused training run with good synthetic data and tight feedback loops could outperform the generalists. If Composer 2.5 is a preview, the next year of AI may look less like one model to rule them all and more like a portfolio of specialists, each cheaper and sharper in its lane than anything a frontier lab ships.
Originally posted on LinkedIn.