Cursor Composer 2 Beats Claude Opus 4.6 at $0.50/1M — Built on Chinese AI

Cursor launched Composer 2 on March 19, 2026. It beats Claude Opus 4.6 on coding benchmarks. It costs $0.50 per million input tokens — one-tenth of what Claude charges. And it was built on top of Kimi K2.5, a Chinese open source model from Moonshot AI.

Moonshot AI is now accusing Cursor of license violation.

That is a lot happening at once. Here is the full breakdown.

What Composer 2 Actually Is

Composer 2 is Cursor's first in-house coding model, replacing the previously used Claude and GPT-5.4 backends for its agentic coding workflows. It is a fine-tuned variant of Kimi K2.5 — Moonshot AI's open source model — with Cursor doing substantial continued training on top.

Cursor employee Lee Robinson confirmed that roughly one quarter of the pretraining comes from the Kimi K2.5 base model, with Cursor responsible for the remaining fine-tuning and continued training. The result is a model purpose-built for the specific task Cursor does: understanding large codebases, making multi-file edits, running terminal commands, and completing long agentic coding tasks.

The Benchmark Numbers

Composer 2 beats Claude Opus 4.6 on Terminal-Bench 2.0 (61.7 vs 58.0) — the benchmark most relevant to what Cursor actually does: running commands, navigating file systems, executing code in real environments. GPT-5.4 still leads at 75.1, but Composer 2 closes the gap substantially.

The improvement over the previous Composer 1.5 is significant: +17 points on CursorBench, +14 points on Terminal-Bench 2.0, +8 points on SWE-bench Multilingual.

The Pricing Disruption

This is where it gets interesting for developers.

Composer 2 Standard is 86% cheaper than Cursor's previous Composer 1.5 ($3.50/$17.50 per 1M). It is 10x cheaper than Claude Opus 4.6 on input and 10x cheaper on output. For developers running high-volume agentic coding tasks, this changes the cost calculation entirely.

A developer running 10 million output tokens per month on Claude Opus 4.6 pays $250. On Composer 2 Standard, that same volume costs $25.

The Kimi K2.5 Controversy

Cursor did not disclose that Composer 2 was built on Kimi K2.5 at launch. The Chinese origin of the base model emerged through technical analysis and subsequent confirmation from a Cursor employee.

Moonshot AI — the Beijing-based company that created Kimi K2.5 — has publicly accused Cursor of license non-compliance. Kimi K2.5 is released under a licence that requires attribution and compliance with specific commercial use terms. Moonshot AI's position is that Cursor fine-tuned and commercialised the model without meeting those requirements.

Cursor has not issued a formal response to the accusation as of March 22, 2026.

This creates an uncomfortable situation for a $29.3 billion American AI coding startup: its flagship coding model, used by one million developers daily, is built on Chinese intellectual property with an unresolved licensing dispute attached to it.

The geopolitical dimension is real. US export controls and AI supply chain scrutiny are intensifying. A major American developer tool depending on a Chinese base model — and then facing accusations of misappropriating that model — is exactly the kind of story that draws regulatory attention.

What Cursor's Scale Looks Like

Cursor now serves more than one million daily active users and 50,000 businesses. Among those businesses: Stripe and Figma. The company is valued at $29.3 billion following its most recent funding round — making it one of the most valuable AI startups outside of OpenAI and Anthropic.

The launch of Composer 2 is Cursor's first step toward model independence. Rather than reselling Claude or GPT-5.4 access, Cursor now has a proprietary model trained on its own data. The strategic logic is clear: control the model, control the margins, control the product roadmap.

What This Means for Developers Using Cursor

If you use Cursor, Composer 2 is available now in the Cursor editor. It runs as the default model for agentic tasks in the Composer interface.

Practically, the benchmark improvements mean Composer 2 handles longer, more complex coding sessions more reliably than Composer 1.5. The Terminal-Bench improvement is the most meaningful for day-to-day use — it measures ability to navigate real development environments, not just generate code snippets in isolation.

The pricing change matters if you are on an API-based Cursor plan. At $0.50/1M input, extended agentic sessions that previously hit token limits become significantly cheaper to run.

The Larger Trend

Cursor is not the only company building on Chinese open source models. The AI supply chain is more intertwined than the public narrative suggests. DeepSeek, Kimi, Qwen, and other Chinese open source models are being used as base layers for products built and sold in Western markets.

Export controls target chips and compute. They do not currently restrict downloading and fine-tuning open source model weights. Until weight controls become policy — which remains politically contentious — Chinese-origin open source models will continue to serve as the foundation for products that never disclose that origin.

The Cursor-Moonshot dispute is the first high-profile case where that foundation has been challenged publicly. It will not be the last.

Key Takeaways

• Cursor Composer 2 launched March 19 — beats Claude Opus 4.6 on Terminal-Bench 2.0 (61.7 vs 58.0), trails GPT-5.4 (75.1)
• $0.50/1M input, $2.50/1M output — 10x cheaper than Claude Opus 4.6, 86% cheaper than Composer 1.5
• Built on Kimi K2.5 — Moonshot AI's Chinese open source model, with Cursor fine-tuning on top
• Moonshot AI is accusing Cursor of license violation — unresolved as of March 22, 2026
• 1 million daily users, 50,000 businesses including Stripe and Figma — Cursor is $29.3B valued
• SWE-bench Multilingual: 73.7 — up from 65.9 on Composer 1.5
• The model independence play — Cursor is reducing dependence on Anthropic and OpenAI APIs by building its own model stack