Vibe Coding

Cursor Launches Composer 2, Its Cheapest Frontier Coding Model Yet

Anysphere's new in-house model undercuts Opus 4.6 and GPT-5.4 on price while matching them on benchmarks.

Andrés Martínez
Andrés MartínezAI Content Writer
March 19, 20262 min read
Share:
Abstract visualization of a coding model processing sequential tasks, with compressed data streams flowing through a neural network

Cursor shipped Composer 2 on Wednesday, the third generation of its in-house coding model and easily the most aggressive on price: $0.50 per million input tokens, $2.50 per million output. For context, Anthropic's Opus 4.6 runs $5/$25 and OpenAI's GPT-5.4 sits at $2.50/$15. A faster variant defaults at $1.50/$7.50, still cheaper than the competition.

The benchmark numbers are strong, at least on paper. Composer 2 scores 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual, up from 47.9 and 65.9 on the previous Composer 1.5. Those are company-reported figures using Cursor's own evaluation harness, so take the exact rankings with a grain of salt. The Terminal-Bench scores for competing models were pulled from official leaderboards or Cursor's internal runs, whichever was higher.

The technical bet here is what Cursor calls self-summarization: during long tasks, the model pauses to compress its own context down to roughly 1,000 tokens, then keeps working. Because this compression happens inside the reinforcement learning loop, the model learns what to keep and what to discard. Cursor says this lets Composer 2 handle tasks spanning hundreds of sequential actions without losing the thread. One demo showed the model working 170 turns to solve a single problem.

Anysphere, Cursor's parent company, now claims over 1 million daily users and 50,000 business customers. Bloomberg reported talks for a new funding round at roughly $50 billion valuation. Composer 2 is available now in Cursor and in an early alpha of a new interface called Glass.


Bottom Line

Composer 2 costs 10x less than Opus 4.6 per input token while posting competitive scores on coding benchmarks, though all numbers are self-reported by Cursor.

Quick Facts

  • $0.50/M input, $2.50/M output tokens (standard); $1.50/$7.50 (fast variant)
  • 61.7 on Terminal-Bench 2.0, 73.7 on SWE-bench Multilingual (company-reported)
  • CursorBench score jumped from 44.2 (Composer 1.5) to 61.3 (Composer 2)
  • Cursor has 1M+ daily users, 50,000 business customers
  • Anysphere reportedly in talks at ~$50B valuation
Tags:CursorComposer 2AnysphereAI codingvibe codingreinforcement learningcoding agents
Andrés Martínez

Andrés Martínez

AI Content Writer

Andrés reports on the AI stories that matter right now. No hype, just clear, daily coverage of the tools, trends, and developments changing industries in real time. He makes the complex feel routine.

Related Articles

Stay Ahead of the AI Curve

Get the latest AI news, reviews, and deals delivered straight to your inbox. Join 100,000+ AI enthusiasts.

By subscribing, you agree to our Privacy Policy. Unsubscribe anytime.

Cursor Launches Composer 2: Frontier Coding at $0.50/M Token | aiHola