Cursor hid Composer 2's base model, but the API exposed Kimi K2.5 and sparked a transparency dispute

The launch of Composer 2 from Cursor looked like a textbook presentation of a new coding model: their own benchmark, strong results, and pricing that undercuts OpenAI and Anthropic. But on March 20, 2026, a single API response destroyed this narrative: hidden under the new name was Kimi K2.5 from Moonshot AI.

How it all came to light

On March 19, Cursor presented Composer 2 as an important step forward for its AI assistant for developers. The company laid the groundwork in advance: eight days before release, it published CursorBench, an internal benchmark for evaluating coding agents on real user requests. The article also described serious engineering work around RL and context compression directly during training. Against the backdrop of neat graphs and aggressive pricing, the launch looked like Cursor didn't just catch up with the leaders, but found a cheaper formula for quality.

On Terminal-Bench 2.0, the new model scored 65.3 points, beating Claude Opus 4.6 with 58.0, although losing to GPT-5.4 with 75.1. The price looked almost provocative: $0.50 per million input tokens and $2.50 per million output tokens, which is 86% cheaper than Composer 1.5.

The problem wasn't in the numbers, but in the silence. In the official launch, Cursor didn't name the base model on which Composer 2 was built, and this exact gap turned out to be critical.

On March 20, a developer named Finn was simply debugging a Cursor endpoint compatible with OpenAI and saw in the response the string `accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast`. This was enough for the entire release packaging to crack within hours.

From the string itself, it was possible to recover almost everything: the Kimi K2.5 model, traces of RL-tuning, the build date, and the internal inference variant. So a technical detail instantly turned into a public investigation into the origin of Composer 2. And without any hacking.

"At least rename the model ID."

Why the dispute started

After the leak, Moonshot AI responded almost immediately. The head of the pretraining direction publicly stated that Composer 2's tokenizer matches Kimi, and suggested that Cursor fine-tuned their model without meeting license conditions. The grounds for the claim weren't unfounded: Kimi K2.5 is distributed under a modified MIT license, which has a separate requirement for large commercial services to prominently display the model name in the interface.

Cursor, according to the article, has monthly revenue significantly above the specified threshold.

• Kimi K2.5's license requires explicitly showing the model name in the product if revenue exceeds $20 million per month or audience exceeds 100 million MAU.
• Cursor, according to estimates in the article, has annual revenue exceeding $2 billion, meaning monthly about $167 million.
• In the interface and marketing post, Cursor didn't mention Kimi K2.5.
• After the scandal, Cursor acknowledged that Composer 2 started with an open-source base, and then directly named Kimi K2.5.

Then the story took a sharp turn. The official Kimi account congratulated Cursor on the release and confirmed that access to Kimi K2.5 went through Fireworks AI as part of an authorized commercial partnership. So the legal basis, apparently, was in order from the beginning.

The main failure happened not in the model and not in the deal, but in communication: part of the Moonshot team accused Cursor publicly, while the official account was already acknowledging the partnership, and Cursor itself took too long to provide a direct explanation.

Why this matters

The most interesting thing about this story isn't that Cursor forgot to rename the model ID. Much more importantly, it looks like a symptom of the entire AI coding industry.

According to the article, Windsurf fine-tunes a version of GLM-4.6 from Zhipu AI, Vercel integrated GLM-4.6 into API services, Cerebras is betting on the same model, and Together AI deployed Qwen-3-Coder from Alibaba.

On OpenRouter, four out of the five most used models in the world are already Chinese, and in early March 2026, their consumption exceeded American analogues for two weeks straight.

Why companies don't like talking about this openly is also clear. First, it's easier to sell investors the story of "our unique model" than of strong fine-tuning of an open-source base. Second, American companies find it awkward to put on the first screen the thesis "runs on Chinese AI." Third, acknowledging the base model shifts focus from the magic of the model itself to things like UX, agent workflow, tools, and RL pipeline. But that's already a less impressive, though more honest conversation about where value is actually created.

What it means

The Composer 2 story shows a simple shift: in AI, it increasingly matters not only which model you took as a base, but also how honestly you talk about it. Cursor really did do notable engineering work, but the attempt to keep the origin of the model off-screen cost it more than a normal acknowledgment of the base from the beginning would have.