A $29B US startup is shipping a Chinese open-source model inside its IDE. They didn't say.
The startup is Cursor.
The model is Composer 2.5.
The base is Kimi K2.5 - Moonshot AI's open-weight model.
Cursor didn't disclose this at launch. A user inspected API headers in March and found it. Cursor confirmed shortly after.
What they actually built: heavy post-training on top of Kimi.
What they actually sell: $0.50 per million input tokens.
What they actually beat: Opus 4.7 on CursorBench at one-tenth the cost per task.
Composer 2.5: ~63% accuracy, $0.50/task.
Opus 4.7: ~62% accuracy, $7/task.
GPT-5.5: ~59% accuracy, $2.20/task.
The actual story of AI coding in 2026:
Frontier labs spend $100M+ training base models.
Moonshot ships them open-weight, for free.
US wrappers post-train, brand, and resell at $0.50/M.
If you've been wondering why Cursor pricing feels suspiciously good, this is why.