Originally published on NextFuture
Cursor shipped Composer 2 in March 2026 as the centerpiece of the Cursor 2.0 overhaul. The model runs on Cursor's own code-only architecture — no Claude or GPT-4 proxy underneath — at $0.50/1M input tokens. The headline number is real, but the KV cache economy behind it is what changes your monthly bill on agentic coding workloads.
What changed
Custom code-only model: Composer 2 is trained exclusively on coding data via continued pre-training and reinforcement learning. Prior Cursor versions routed all Composer requests to Claude Sonnet or GPT-4 through the Anthropic and OpenAI APIs.
$0.50/1M input tokens: That's 6× cheaper than Claude Sonnet 3.7 ($3/1M) and 16× cheaper than GPT-4o ($8/1M) — the two models Cursor previously proxied for Composer tasks.
Aggressive KV cache on repo context: Cursor stores repo embeddings in a shared KV cache between agentic requests. Follow-up edits to the same files in one session read cached tokens at under $0.05/1M instead of re-ingesting the full context.
SWE-bench Multilingual leadership: Composer 2 outperforms GPT-4o and Claude Sonnet 3.7 on this benchmark, which tests multi-file edits across Python, Java, Go, Rust, and TypeScript — not just synthetic toy problems.
General reasoning dropped by design: Non-coding knowledge was excised during training. The model is intentionally limited outside code — ask it about system architecture or documentation and it will redirect you to a general-purpose model.
Why builders should care
At 1,000 Composer requests per day — a realistic number for any team running overnight refactors or CI-triggered context expansions — moving from Claude Sonnet at $3/1M to Composer 2 at $0.50/1M cuts input costs by 83%. Teams that spent $90/month on proxied frontier model tokens in 2025 can run the same agentic coding volume for roughly $15/month. That delta matters most for solo operators and small teams without enterprise AI budgets who are paying out-of-pocket for agentic coding pipelines.
The KV cache amplifies the savings further. In a long session where Cursor reuses cached repo context across 20–30 sequential edits, effective per-token cost on the cache-hit portion drops below $0.10/1M. The practical result: a codebase-wide refactor that cost $4 in tokens last year costs under $0.50 today. Cost-sensitive workloads — background agents, automated PR reviews, multi-step migration scripts — benefit most.
What changes in your workflow
Model routing is automatic from Cursor 2.0+: Composer defaults to Composer 2. If you previously hardcoded
claude-sonnet-3-7orgpt-4oin the Cursor model selector, reset it to Auto or explicitly choose Cursor Composer 2 in Settings → AI → Composer Model.No API key changes for standard subscribers: Composer 2 is internal to the IDE on Cursor's Pro and Business plans. If you use Cursor's API for CI integrations, verify the
cursor.composer.modelfield in your project config points tocomposer-2.Cache read tokens appear separately in billing: The Cursor billing dashboard now distinguishes "cache read" from "input" tokens. Expect cache reads to dominate the token breakdown in sessions that edit the same files repeatedly.
Route non-code questions to Cursor Chat, not Composer: Composer 2 is intentionally blunt on architecture discussions and documentation. Use the full-context Chat models (Claude or GPT-4o, selectable in the chat panel) for anything requiring general reasoning.
Check your Cursor version and confirm the model selector with:
# Confirm you're on Cursor 2.0+
cursor --version
# → Cursor 2.0.x (Composer 2 is the default Composer model from 2.0 onward)
# List available Composer models via CLI (Cursor 2.0+)
cursor models --composer
5 action items for this week
Update Cursor to 2.0 or later. Run
cursor --versionin a terminal. If you're on 1.x, download the latest from cursor.sh — Composer 2 ships only with 2.0+.Switch Composer's model to "Cursor Composer 2" in Settings → AI → Composer Model. "Auto" already routes there on 2.0, but explicit selection ensures you don't fall back to a proxied frontier model on timeout.
Check your Cursor billing dashboard at cursor.sh/billing. Look at the per-model token breakdown. If you see a large "Claude Sonnet" input line still accumulating, your Composer is not on 2.0 yet — update and reverify.
Keep Claude or GPT-4o active in Cursor Chat (not Composer) for architecture discussions, inline documentation, and general Q&A. Set your preferred chat model in Settings → AI → Chat Model so the fallback is explicit.
Benchmark your heaviest refactor case. Run Composer 2 on the most context-intensive multi-file task you do regularly and compare output quality and token cost against your Claude Sonnet baseline. SWE-bench Multilingual predicts it will win on edit-heavy tasks; verify it holds on your actual codebase.
What to watch next
Cursor's move to a proprietary code-only model puts direct pricing pressure on every AI coding tool that still proxies frontier models. The $0.50/1M vs $3–8/1M gap is now a hard TCO argument in enterprise procurement conversations. If your team is actively evaluating which AI editor to standardize on, see our 10 Best Cursor Alternatives in 2026 roundup — several tools there still rely on OpenAI or Anthropic proxies, and the cost comparison looks different today than it did six months ago.
The RL-from-code-feedback methodology Cursor uses is also worth tracking. Specialized models trained on real bug-fix and PR-diff data are consistently outperforming general-purpose frontier models on coding benchmarks. For a broader view of where agentic coding tooling is headed, the Claude Code /advisor deep dive covers strategic planning layers on top of code-only execution.
FAQ
Is Cursor Composer 2 cheaper than routing through Claude Sonnet via the Cursor API?
Yes, by a factor of 6×. Composer 2 costs $0.50/1M input tokens. Claude Sonnet 3.7 — which Cursor previously proxied — runs at $3/1M input via Anthropic's API. At 10M tokens/month, that's $5 vs $30 in input costs alone, before counting cache-read savings that push Composer 2's effective cost lower still in long sessions.
Does Cursor Composer 2 work well with non-TypeScript codebases?
Yes. The SWE-bench Multilingual benchmark that Composer 2 leads on explicitly tests Python, Java, Go, Rust, and C++ alongside TypeScript. The continued pre-training corpus spans multiple languages, and the RL phase trains on real multi-language edit data from open-source repositories. Performance on TypeScript/Next.js will likely remain the highest-tested case, but Python and Go codebases should see competitive results.
When should I use Cursor Chat instead of Cursor Composer 2?
Use Composer 2 for any task that produces code: file edits, multi-file refactors, test generation, and migration scripts. Switch to Cursor Chat (Claude or GPT-4o) for architecture decisions, writing documentation, explaining third-party library internals, or any question where general reasoning matters. Composer 2 is trained to produce correct code edits, not to reason about system design — the model will redirect you if you try to use it outside that scope.
This article was originally published on NextFuture. Follow us for more fullstack & AI engineering content.
Top comments (0)