Cursor Composer 2: Pricing, Benchmarks & Kimi K2.5 Base

The third Composer in five months

Cursor's announcement sets two prices: a standard variant at $0.50 per million input tokens and $2.50 per million output tokens, and a faster variant at $1.50 and $7.50 for the same intelligence at lower latency. Cursor's own benchmark table shows the progression across its three models: Composer 1 scored 38.0 on CursorBench, 40.0 on Terminal-Bench 2.0 and 56.9 on SWE-bench Multilingual; Composer 1.5 improved that to 44.2, 47.9 and 65.9; Composer 2 reaches 61.3, 61.7 and 73.7. [1]

Cursor's technical report describes a two-phase process behind that jump: what it calls its first continued pretraining run to build coding knowledge into the base weights, followed by large-scale reinforcement learning on long-horizon coding tasks, using training infrastructure built to match Cursor's production environment. [2]

Atlas interpretation: Composer 1, 1.5 and 2 in roughly five months is a release cadence closer to a point update than a model line, and it reframes what the benchmark table is actually showing: not a single model's strength, but how fast Anysphere can turn training runs into a shipped price and a shipped score. [1][2]

An undisclosed open-weight base, found by a user on X

Cursor's own announcement and technical report describe Composer 2 as trained from Cursor's own pipeline, without naming an outside base model. Three days after launch, an X user identified as Fynn reported finding code markers showing Composer 2 was actually built on top of Kimi K2.5, the open-weight trillion-parameter model Moonshot AI had released seven weeks earlier, and posted the finding publicly. [3]

Cursor VP Lee Robinson confirmed it afterward, writing that Composer 2 started from an open-source base and that only about a quarter of the final compute came from that base, with the rest from Cursor's own training. Co-founder Aman Sanger went further: "It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model." Robinson also said the arrangement ran through a licensing deal with Fireworks AI, and that Composer 2's benchmark behavior had diverged substantially from Moonshot AI's original checkpoint despite the shared starting point. [3]

Atlas interpretation: Nothing about using a licensed open-weight base is against Cursor's own rules, and Robinson's compute math is the kind of detail a company volunteers when it wants a story to end. What made the omission notable was the source: an American company building a flagship coding model on top of a Chinese lab's checkpoint, in an industry still replaying its DeepSeek anxieties, and choosing not to say so until a user on X worked it out first. [3]

Two months later, Cursor said it upfront

Cursor's announcement of Composer 2.5 states plainly that it "is built on the same open-source checkpoint as Composer 2, Moonshot's Kimi K2.5," and describes the newer model as differing mainly in post-training: roughly 25 times more synthetic training tasks and additional reinforcement learning, rather than a new base. [4]

Atlas interpretation: Sanger's promise to fix the disclosure next time held. Composer 2's base came out sideways, through a stranger's read of the model's own code output; Composer 2.5's came out in the announcement's second paragraph. [4]