Top AI Product

Cursor shipped Composer 2.5 on May 18 — an in-house coding agent built on the open-source Kimi K2.5 checkpoint from Moonshot AI, then heavily post-trained by Cursor (roughly 85% of total compute budget went into Cursor’s own reinforcement learning and post-training pipeline). The headline: 79.8% on SWE-Bench Multilingual, matching Claude Opus 4.7 and GPT-5.5 at roughly one tenth the cost per token.

## The pricing

Standard tier: $0.50 per million input tokens, $2.50 per million output tokens. Fast tier: $3.00 in, $15.00 out. The fast pricing already undercuts most frontier closed-source models, while standard pricing is aggressive enough to change unit economics for agent-heavy workflows where you burn through tens of millions of tokens per session.

## What improved over Composer 2

25x more synthetic task training, refined behavioral calibration, and substantial gains on long-running tool-heavy sessions — the kind where an agent reads files, runs terminal commands, edits across multiple files, executes tests, and iterates rather than answering a single one-off prompt.

## Why it matters

Cursor announced they’re training a significantly larger model from scratch in collaboration with xAI, using Colossus 2’s million H100-equivalents and 10x more total compute. Composer 2.5 is the proof-of-pipeline before that bigger bet lands. If Cursor can hit frontier benchmarks at 10% of frontier prices today, the AI coding race shifts from model competition to vertical integration of editor and model.