Cursor Composer 2 just beat Claude Opus — at 86% less cost

A Coding Model That Costs Less and Scores Higher

On March 19, 2026, the AI coding platform Cursor quietly released something that raised eyebrows across the developer community: a new model called Composer 2 that beats Anthropic's Claude Opus 4.6 on coding performance tests — while costing 86% less per query.

Cursor is an AI-powered code editor (a program that helps you write software faster by suggesting, editing, and even writing entire files for you) with a reported valuation of $29.3 billion. Composer 2 is its latest in-house model, designed specifically for coding tasks.

What the Benchmarks Actually Show

To measure how well AI coding tools perform, researchers use benchmarks (standardized tests that challenge the AI with real coding problems so different models can be fairly compared). Composer 2 was tested on three of these:

• CursorBench: Composer 2 scores 61.3 vs. Claude Opus 4.6 at 58.2 — Cursor's own internal test suite measuring complex editing tasks
• Terminal-Bench 2.0: Composer 2 scores 61.7 vs. Claude Opus 4.6 at 58.0 — a test of real-world software tasks carried out inside a terminal (the text-based command interface on a computer)
• SWE-bench Multilingual: Composer 2 scores 73.7 — measuring how well the model can resolve actual bugs across multiple programming languages

For context, those same scores show a significant leap over Composer 2's predecessor: Composer 1.5 scored just 44.2 on CursorBench and 47.9 on Terminal-Bench. The new model improved by nearly 40% on those same tests.

GPT-5.4 from OpenAI still leads the field at 75.1 on Terminal-Bench 2.0, but Composer 2 now sits clearly ahead of Claude Opus 4.6, which was previously considered the gold standard for difficult coding tasks.

The Price Gap Is Hard to Ignore

AI models charge by the token (a unit of text roughly equal to three-quarters of a word — so 1,000 tokens is about 750 words). Here's how the pricing stacks up:

• Composer 2 Standard: $0.50 per million input tokens / $2.50 per million output tokens
• Composer 2 Fast (same accuracy, faster response): $1.50 input / $7.50 output per million tokens
• Claude Opus 4.6: $5.00 input / $25.00 output per million tokens
• GPT-5.4: $2.50 input / $15.00+ output per million tokens

Put simply: Composer 2 Standard costs 90% less than Claude Opus 4.6 per input token. Even the faster, premium variant of Composer 2 is 70% cheaper than Opus. For a team generating 10 million output tokens per month, that's the difference between a $250 monthly bill (Opus) and a $25 monthly bill (Composer 2 Standard).

The 86% figure comes from comparing Composer 2 to its predecessor Composer 1.5, which was priced at $3.50 per million input tokens. The new model drops that to $0.50 — a dramatic cut that Cursor says reflects both efficiency improvements and its strategy to compete on price.

How Cursor Built It — and the Controversy That Followed

Within hours of the Composer 2 launch, a developer named Fynn intercepted traffic from the Cursor app and discovered the underlying model identifier: kimi-k2p5-rl-0317-s515-fast. This revealed that Composer 2 is built on top of Kimi K2.5, an open-source model from Moonshot AI, a Chinese startup.

Cursor had not disclosed this in its launch blog post, which sparked a pointed debate on Hacker News (a popular forum where developers and engineers discuss technology news). Critics argued that Cursor — valued at nearly $30 billion — was primarily packaging and reselling open-source work rather than building original AI from scratch.

Cursor co-founder Lee Robinson later confirmed the foundation and clarified: roughly 25% of the compute used to build Composer 2 came from the Kimi K2.5 base, with the remaining 75% from Cursor's own continued pretraining (additional training on more data after the base model is built) and reinforcement learning (a training method where the AI learns by receiving rewards for correct answers, similar to how you'd teach a dog new tricks).

Moonshot AI confirmed the arrangement was licensed through Fireworks AI, an authorized commercial partnership, settling concerns about legal compliance.

What's Actually New in Composer 2

Beyond the benchmark numbers, Cursor highlights several technical capabilities that make Composer 2 better for long, complex coding jobs:

• 200,000-token context window: The model can read and process roughly 150,000 words at once — enough to hold an entire large codebase in its working memory at the same time
• Hundreds of sequential actions: Composer 2 can carry out multi-step coding jobs that require making dozens or hundreds of changes across many files without losing track
• Self-summarization: For very long tasks, the model summarizes its own progress so it doesn't forget what it was doing at the start — a key limitation of earlier AI tools
• Tool use and file editing: The model is tuned specifically for using developer tools, editing files, and running terminal commands — not just answering questions

How to Access Composer 2 Right Now

Composer 2 is available directly inside the Cursor editor. The Fast variant is set as the default for Cursor users. To try it:

# 1. Download or update Cursor from:
#    https://cursor.com

# 2. Open any project and press:
#    Ctrl+I (Windows/Linux) or Cmd+I (Mac)
#    to open the Composer panel

# 3. In the model dropdown, select:
#    Composer 2 or Composer 2 Fast

# 4. Composer usage on individual plans
#    falls within included credits — no extra charge to start

Cursor also has an early-access interface called Glass where Composer 2 is being tested in a new environment. You can request access at cursor.com.

For developers accessing the model via Cursor's API (a way for software programs to talk to each other directly), the standard pricing of $0.50/$2.50 per million tokens applies, making it one of the most cost-effective frontier coding models available today.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News