On May 18, 2026, Cursor quietly dropped something significant: Composer 2.5, its own proprietary coding agent model built directly into the Cursor IDE. This isn't a wrapper around an existing frontier model — it's a purpose-built agent trained from the ground up to handle the messy, multi-file, real-world tasks that trip up general-purpose LLMs. If you're a developer living inside Cursor, here's everything you need to know.
What Is Composer 2.5, Exactly?
Composer 2.5 is Cursor's in-house coding agent — not to be confused with PHP's dependency manager of the same name. It operates as the autonomous "agent mode" inside the Cursor IDE, capable of planning, editing, and verifying changes across multiple files in a single run.
The model is built on Moonshot's open-source Kimi K2.5 base checkpoint, but Cursor made a decisive architectural bet: dedicating 85% of the total compute budget to post-training and reinforcement learning rather than pre-training. The result is a model whose raw intelligence comes from Kimi K2.5, but whose coding behavior is almost entirely Cursor's own. (Lushbinary)
The Training Methodology: Why It Matters
Most coding models are evaluated on whether they produce correct output at the end of a task. Cursor took a different approach with Composer 2.5, and the details are worth understanding.
25x More Synthetic Training Tasks
Composer 2.5 was trained on 25 times more synthetic tasks than Composer 2. One standout task type is "feature deletion" — the model is given a codebase with a feature removed and must reimplement it from scratch while keeping all tests green. This forces the agent to reason about intent, not just syntax. (Cursor Blog)
Targeted Reinforcement Learning with Textual Feedback
Rather than rewarding the model only at the end of a successful run, Cursor implemented localized natural-language hints at each failed tool call. When the agent makes a wrong move — a bad file edit, a broken import — it receives specific textual feedback in the moment, not just a binary pass/fail at the finish line. (OfoxAI)
This mid-trajectory feedback loop is a meaningful departure from standard RLHF pipelines and likely explains much of the model's strong performance on long-horizon agentic tasks.
Benchmark Performance: How It Stacks Up
Cursor published results across three benchmarks. The numbers position Composer 2.5 squarely in frontier territory — and ahead of some well-known competitors on key tests.
CursorBench v3.1: Composer 2.5 scored 63.2%, outperforming Claude Opus 4.7 (61.6%) and GPT-5.5 (59.2%) on Cursor's own internal benchmark suite. (OfoxAI)
SWE-bench Multilingual: The model achieved 79.8%, just behind Claude Opus 4.7's 80.5% but ahead of GPT-5.5's 77.8% — a strong result on a third-party, language-diverse benchmark. (OfoxAI)
Artificial Analysis Coding Agent Index: Composer 2.5 scored 62, a 14-point jump over Composer 2's score of 48 — the largest generational leap in the model's history. (Totalum)
Important: CursorBench v3.1 is Cursor's own benchmark, which means the company controls both the model and the evaluation. Weight the SWE-bench Multilingual and Artificial Analysis scores more heavily when comparing against external models.
Pricing and Access
Composer 2.5 runs exclusively within the Cursor IDE and is not available as a standalone API. Developers cannot call it from external tooling or pipelines — it is a Cursor-native product. (DataCamp)
Pricing is tiered by speed, giving teams flexibility based on their workflow needs. (US Tech Automations)
Standard tier: $0.50 per million input tokens and $2.50 per million output tokens — suited for exploratory work, refactoring sessions, and non-time-sensitive tasks.
Fast tier: $3.00 per million input tokens and $15.00 per million output tokens — designed for latency-sensitive workflows where response speed directly impacts developer productivity.
Pro Tip: For most day-to-day agentic tasks — feature implementation, test generation, bug triage — the Standard tier will deliver the same model quality at a fraction of the cost. Reserve the Fast tier for live pair-programming sessions or time-boxed sprints.
Key Takeaways
Purpose-built for agentic coding: Composer 2.5 is not a general-purpose LLM — it is trained specifically for multi-file, long-horizon coding tasks inside a real IDE environment.
Novel training approach: 85% of compute went to post-training and RL, with mid-trajectory textual feedback replacing blunt end-of-run rewards.
Competitive benchmark results: Scores of 63.2% on CursorBench v3.1 and 79.8% on SWE-bench Multilingual place it alongside — and sometimes ahead of — frontier models from Anthropic and OpenAI.
Cursor-only access: There is no external API. If you want Composer 2.5, you work inside Cursor — full stop.
Flexible pricing: A Standard tier at $0.50/$2.50 per million tokens makes the model accessible; the Fast tier at $3.00/$15.00 targets teams where latency is a competitive factor.
