> Source: https://cursor.com/blog/router > Rating: 70 · Category: ai-products · Published: 2026-07-22
On July 22, Cursor released Cursor Router. It is not a new model—it's an intelligent routing system. Cursor classifies all user requests and routes them to the most appropriate model at the moment. It sounds mundane, but it marks a key step: Cursor turning "model selection" from a user decision into a product decision.
Cursor's data tells one story: about 60% of developers use a single model as their "daily driver." The result is that even routine tasks get billed at frontier prices, and AI spend grows faster than output quality. That is exactly the problem Cursor Router is built to solve.
The numbers
- Early access: 3 high-usage enterprise accounts (thousands of users) saved 30–50% under Auto routing compared to running everything on Opus 4.8 API pricing—with no quality degradation.
- Online A/B testing (millions of requests): Cursor Router achieved frontier-level user satisfaction with 60% cost savings.
- Auto Intelligence mode: satisfaction close to Fable 5 at about 60% of Fable 5's cost; at equal cost, satisfaction roughly 15% higher than Opus 4.8.
- Auto Balance mode: satisfaction exceeds Opus 4.8 with ~36% savings; vs GPT-5.6 Sol, comparable satisfaction at lower spend.
- Cost per commit: Intelligence mode $6.76; Balance mode $4.63; GPT-5.6 Sol matches Intelligence on cost with lower satisfaction; Fable 5 is $12.69; Opus 4.8 is $7.34.
- Training data: 600,000 real online requests.
- Evaluation: online A/B testing, comparing Cursor Router's decisions against "always use the same model."
- Optimization target: user satisfaction (Agent Feedback Comment, AFC) — continuing to the next task = strong positive signal; correcting the agent = strong negative signal. There's also a Keep rate: how much agent-generated code ultimately survives in the codebase.
- Intelligence: frontier performance, even at higher cost—push the ceiling.
- Balance: the cost-performance sweet spot; most people's daily driver.
- Cost: minimize token budget under acceptable quality.
- Lessons for other vendors: Cursor's "model neutrality" has prerequisites—it owns Composer and gets real traffic data from all frontier models for training. Anyone building a similar router first needs to solve data access and model-pool problems.
- "Auto Intelligence ≈ Fable 5 satisfaction at 60% cost" is the number to watch. Fable 5 (essentially an upgraded Claude Opus 4.8) is the strongest Anthropic model, yet a routed ensemble nearly matched its satisfaction. This again validates the July 19 Schema Harness takeaway: "keep the model, change the harness" is the decisive play for 2026 H2.
- Per-commit cost comparison: $4.63 (Balance) vs Fable 5's $12.69 makes "should we switch the team from Fable 5 to Cursor Router" a straightforward engineering decision.
- July 16: Anthropic self-migrated 1M lines from Zig to Rust with Bun in 11 days, at a cost of $165,000.
- July 19: Cursor ran GPT-5.5 Composer + Opus 4.8 planner + Grok 4.5 executor for 4 hours to reach 80% SQLite coverage, costing $411.
- July 19: Schema Harness scored 99% on the ARC-AGI-3 public set—"keep the model, change the harness."
- July 22: Cursor Router ships model selection as a product-level default.
By putting these numbers front and center, Cursor is essentially selling engineering-team leads a cost-effectiveness story.
How it decides which model to call
Cursor Router is a classifier: for each request, before and after dispatch, it examines the query, context, task complexity, and domain, combined with Cursor's known per-model characteristics. Simple work goes to the cheapest model, UI changes go to the model with the best "taste," and complex long-horizon reasoning goes to a frontier reasoning model.
One critical design detail: the classifier accounts for cache hits in both training and evaluation. Switching models triggers a cache miss, and the router does not pretend otherwise. In other words, the cost-savings numbers reflect real-world conditions, not optimistic offline benchmarks.
Three modes + team governance
Auto mode offers three tiers:
Administrators can enable Cursor Router at team/group granularity, decide which modes members may choose, set defaults, and block specific models. For enterprise IT, that's a rare level of governance granularity for an "AI coding tool."
Engineering details worth unpacking
First, Cursor Router is cache-aware—and that's hard. Many "model routing" prototypes demo well but never ship because they assume switching models is free. Cursor built it in as a training and evaluation constraint, not an afterthought.
Second, the model pool is expanding in both directions. Cursor linked the July 19 release of Grok 4.5 the same day—"expanding Cursor Router's options for hard, expensive tasks"—while pointing to its own Composer models below—"daily tasks approaching frontier quality, where the savings are most visible." Pulling on both ends means Cursor now has a two-layer pool: its own base models plus external frontier models.
Third, Cursor Router is one piece of Cursor's broader "token efficiency" puzzle. The blog explicitly lists it alongside Dynamic Tool Calling and the Continually Improving Agent Harness. Dynamic Tool Calling keeps most tool descriptions out of the prompt, letting models look them up on demand—the same pattern as MCP. Cursor's message: model selection is just the entry point; the harness itself must keep slimming down.
Some hooks
Timeline: one week of AI coding tool evolution
Source: Cursor official blog, 2026-07-22, "Introducing Cursor Router," by the Cursor team.