Cursor Composer 2.5 nears Opus 4.7 and GPT-5.5 at a fraction of the cost

Cursor released Composer 2.5 on May 18, its latest in-house coding model, and the pitch is not that it beats the frontier. It is that it gets close enough to Claude Opus 4.7 and GPT-5.5 on real coding benchmarks while costing a fraction as much to run, which changes the math for any engineering team choosing a default model.

Third-party benchmarking from Artificial Analysis backs the core claim: on the Coding Agent Index, Composer 2.5 lands close behind high-effort configurations of Opus 4.7 and GPT-5.5, but at a small fraction of the per-task cost. For teams evaluating AI coding tools against an actual budget, the token pricing is the more decision-useful number than the benchmark score alone.

What is actually new

Composer 2.5 is built on the same open-weight base as its predecessor, Moonshot’s Kimi K2.5, retrained with more complex reinforcement-learning environments and a new method for assigning credit across long task rollouts. Cursor says the result is meaningfully better at sustained, long-running work and at following complex, multi-step instructions without losing the thread.

Learn more about Composer 2.5: https://t.co/du1AwrYVNB

— Cursor (@cursor_ai) May 18, 2026

Here is the token pricing that matters for budgeting, per million tokens:

At standard rates, Composer 2.5 runs at roughly a tenth of Opus 4.7’s input cost and a twelfth of GPT-5.5’s. Artificial Analysis’s task-level testing found the gap widens further in practice, since Composer 2.5 also used fewer tokens to reach comparable results, up to 20 times cheaper per task than max-effort Opus 4.7 on their coding agent benchmark.

What this means for engineering budgets

The benchmark gap between Composer 2.5 and the frontier models is real but narrow: within a few points on SWE-Bench Multilingual and CursorBench, and GPT-5.5 still leads clearly on Terminal-Bench, the test most relevant to complex, multi-step DevOps-style tasks. That is the practical decision framework for a team setting model defaults: route routine implementation, bug fixes, and boilerplate through the cheap model, and reserve Opus 4.7 or GPT-5.5 for the planning and edge-case work where the accuracy gap actually shows up.

This is the same discipline we described in margin, not growth, is the real scoreboard in AI software: the vendors winning on unit economics are the ones treating model choice as a routing problem, not a single default setting. It also reinforces a point from our buyer’s guide to agentic AI lock-in, that platforms able to route across multiple model providers hold more leverage than ones locked to a single frontier model’s pricing.

What to watch

Cursor has said it is training a substantially larger model from scratch with roughly 10 times more compute, in partnership with SpaceX’s AI efforts. If that model closes the remaining benchmark gap with Opus 4.7 and GPT-5.5 while keeping anything close to current pricing, it would put real pressure on how OpenAI and Anthropic price their own coding-focused tiers. And for where Composer’s current foundations actually come from, see our explainer on Kimi, the Chinese open-weight model it is built on.