Cursor launched Composer 2 on March 20, positioning its AI coding model as a direct challenger to OpenAI and Anthropic at a fraction of the cost. Priced at $0.50 per million input tokens, Composer 2 undercuts frontier models by 80% while beating Claude Opus 4.6 on coding benchmarks. The timing aligns with Cursor’s talks to raise funding at a $50 billion valuation, up from $29.3 billion four months ago. With 1 million daily users and 50,000 business customers including Stripe and Figma, the San Francisco startup is forcing the AI coding market to compete on both price and performance.
Composer 2 Beats Opus 4.6 on Coding Benchmarks
Composer 2 scores 61.7 on Terminal-Bench 2.0, surpassing Claude Opus 4.6’s 58.0 and Opus 4.5’s 52.1. GPT-5.4 leads at 75.1, but Composer 2 delivers competitive performance at a fraction of the cost. On CursorBench, which measures solution correctness and code quality, Composer 2 achieves 61.3. The model also posts 73.7 on SWE-bench Multilingual, covering 300 tasks across nine programming languages.
Terminal-Bench 2.0, maintained by the Laude Institute, tests an AI agent’s ability to perform real-world terminal tasks: navigating directories, running scripts, interpreting errors, and iterating toward solutions. These are actual developer workflows, not synthetic benchmarks designed to flatter specific models.
Aggressive Pricing: $0.50 Per Million Tokens
Cursor offers Composer 2 in two tiers. Standard costs $0.50 per million input tokens and $2.50 per million output tokens. Fast (the default) runs $1.50 input and $7.50 output per million tokens. Both undercut typical frontier model pricing of $2.50 to $5.00 input and $10 to $15 output per million tokens. Bloomberg reports Cursor hit $2 billion in annualized revenue by February 2026, validating the pricing strategy at scale.
Kimi K2.5 Base Model Sparks Attribution Debate
Cursor initially described Composer 2 as an in-house model built through continued pretraining and reinforcement learning. Within 24 hours, developer Fynn discovered an internal model identifier revealing Composer 2 started from Kimi K2.5, an open-weight model from Moonshot AI.
Cursor VP Lee Robinson clarified that only 25% of the compute came from the base model, with 75% from Cursor’s training. Moonshot AI confirmed an authorized commercial partnership through Fireworks AI. Kimi K2.5’s Modified MIT License requires products with over 100 million monthly active users or $20 million in monthly revenue to display “Kimi K2.5” prominently. Cursor failed to provide this attribution at launch.
The controversy highlights transparency expectations in AI development. Fine-tuning open-source models is standard practice, but developers expect clear disclosure about base models. Cursor’s 75% compute contribution represents genuine work, but the attribution misstep damaged trust unnecessarily.
Technical Edge for Long-Horizon Coding Tasks
Composer 2 uses a Mixture-of-Experts architecture where only a subset of parameters activate per input, keeping inference fast. The model offers a 200,000-token context window and is trained exclusively on code, reducing hallucinations for software engineering tasks.
The key innovation is “self-summarization,” a training technique that reduces compaction error by 50% and achieves summaries 5x more token-efficient than standard approaches. This optimizes the model for multi-file refactoring, long task chains spanning hundreds of actions, and terminal operations where context accumulates across many steps.
AI Coding Market Reaches Maturity
The 2025 Stack Overflow Developer Survey shows 84% of developers use or plan to use AI coding tools, with 51% using them daily. By 2026, AI tools write or assist 41% of all code. GitHub reports over 51% of commits in early 2026 were AI-generated or AI-assisted. The AI coding tools market reached $12.8 billion in 2026, up from $5.1 billion in 2024.
Cursor’s growth reflects this trend. The company serves 1 million daily users and 50,000 businesses. Its $50 billion valuation talks signal that specialized AI coding tools are now core infrastructure, not experimental features. The market is shifting from general-purpose LLMs with coding capabilities to purpose-built models optimized for specific workflows.
Composer 2 represents this specialization. Rather than competing across all language tasks, it focuses exclusively on code—particularly the long-horizon, multi-file tasks where general models struggle with context and coherence. As the AI coding market matures, expect more vertical specialization: models for backend systems programming, frontend frameworks, data pipelines, and infrastructure automation, each optimized for domain-specific patterns.
