Cursor ships Composer 2.5 — internal long-horizon model matches Opus 4.7 and GPT-5.5 on benchmarks at one tenth the per-token cost
Cursor shipped Composer 2.5 on May 18, an in-house long-horizon model that matches Anthropic Claude Opus 4.7 and OpenAI GPT-5.5 on two of three public coding benchmarks — at $0.50 per million input tokens and $2.50 per million output tokens. That's approximately one tenth the per-token cost of the closed-frontier alternatives, and Cursor controls the entire stack from training to deployment.
The pricing inflection is the headline, but the architectural fact is what makes it durable. Cursor is the first developer-tools company to ship a frontier-class model trained explicitly on its own user workload — not a fine-tune over a foundation model, but a from-scratch internal training run optimized for the long-horizon coding tasks the company sees from its 2+ million weekly users. Composer 2.5 isn't trying to be a general-purpose frontier model; it's trying to be the best coding model, and on benchmarks where coding is the explicit task, it matches the closed frontier at one tenth the cost.
The implication for the closed-frontier vendor pricing: Anthropic Claude Opus 4.7 and OpenAI GPT-5.5 both list at roughly $15-30 per million output tokens for top-tier access. If Cursor's $2.50/M output mark holds and the coding-task benchmark parity is reproducible by independent reviewers, the entire developer-tools market gets a credible price floor that's an order of magnitude below the closed-frontier list price. That repricing pressure compounds into the JATF / USSOCOM-class procurement conversations Niles is having: "4-bit TriSeq quantization for $50/airframe edge silicon" is now competing not against Anthropic at $30/M tokens but against Cursor at $2.50/M tokens, which closes the cost-defensibility gap by an order of magnitude.
Lushbinary — Composer 2.5 vs Opus 4.7 vs GPT-5.5 Coding Comparison → · WhatLLM — New AI Models May 2026 The Frontier Took a Breath → · AI Thority — From GPT-5.5 to DeepSeek V4 Building Smarter AI Agents →