Cursor Router becomes generally available for coding model selection
Pinning one frontier model for every coding turn just got harder to defend. Cursor shipped Cursor Router to general availability: Auto mode now classifies each request and sends hard work to frontier models while price-efficient models handle the rest. Per the company changelog and announcement, the bet is frontier-grade output without frontier rates on every refactor.
That answers a budget problem engineering leaders already feel. Roughly 60% of Cursor developers, Cursor says, still pick a single daily driver. Routine edits then bill like long-horizon reasoning.
How the request-level classifier picks a model
Cursor Router is a classifier, not a preference toggle. Cursor says it trained on more than 600,000 live requests and scored online A/B tests across millions of production turns, optimizing for user satisfaction (AFC) rather than offline rubrics. Each request is scored on query, context, task complexity, and domain, then matched to what Cursor has learned about model behavior.
Simple work goes to cheaper models. UI polish goes to models with better “taste.” Long-horizon reasoning stays on frontier systems. The router is cache-aware: reported savings include the cache-miss cost of switching models mid-conversation, which offline evals usually ignore. Grok 4.5 is required as a price-efficient routing option; block too many underlying models and routing quality drops.
Cost vs Balance vs Intelligence modes
Admins and users pick where they sit on the cost-intelligence curve:
- Intelligence: frontier quality for the hardest tasks, without locking every turn to the most expensive model.
- Balance: strong quality aimed at the models most teams already daily-drive.
- Cost: good quality while optimizing token spend (Cursor describes this as the previous Auto path with bundled Auto pricing).
Balance and Intelligence bill at the routed model’s rate. Cost keeps bundled Auto pricing. Router runs across desktop, web, iOS, CLI, and the SDK. It is on by default for Teams; Enterprise admins enable it and can restrict modes, set defaults, and allow or block models. Soft and hard enforcement options exist for standardizing on Auto. The routed model can be shown or hidden (hidden by default).
What online A/B tests claimed on satisfaction
Vendor numbers, labeled as such. Cursor says Auto Intelligence lands near Anthropic’s Fable on satisfaction at about 60% lower team cost, and lifts satisfaction about 15% over Opus 4.8 at nearly the same spend. Auto Balance, Cursor claims, beats Opus 4.8 on satisfaction at about 36% lower cost, and matches GPT-5.6 Sol satisfaction at a lower spend rate.
Quality proxies were user satisfaction (move on vs correct the agent) and keep rate (how much generated code stays in the repo). Treat the A/B story as Cursor’s evidence pack, not an independent audit.
Enterprise spend savings vs pinned frontier models
Early access covered dozens of enterprises. Cursor compared what customers paid against the same traffic priced entirely at Opus 4.8 API rates. Three high-volume accounts with thousands of users, Cursor says, saved 30%–50% on Auto-routed requests versus pinning Opus 4.8, with no quality drop in its metrics.
Cost per commit is the sharper enterprise line: Cursor reports $6.76 for Intelligence and $4.63 for Balance, versus $12.69 for Fable 5 and $7.34 for Opus 4.8. GPT-5.6 Sol matched Intelligence on cost but lagged on satisfaction, per Cursor.
Analysis: model routing is how IDE vendors turn neutrality into a margin story. If your shop still staples every agent turn to one frontier SKU, Router makes that habit look like waste. If you need a named model for compliance or evals, keep pinning. For everyone else, the fight shifts from “which model?” to “who controls the classifier, and can we trust the savings after cache misses?”
