Cursor Router: Automatic Model Selection for Auto Mode
CursorCursor introduced Cursor Router, an intelligent classifier that now powers Auto mode by automatically routing every request to the model best suited for the task. Trained on more than 600,000 live coding requests, the router weighs query, context, task complexity, and domain to send simple work to cost-efficient models and demanding, long-horizon problems to frontier reasoning models. Users can pick from three optimization modes, Intelligence, Balance, and Cost, and Cursor reports 30-60% lower costs than using premium models exclusively, with no drop in output quality or code retention. The feature is live now across desktop, web, iOS, CLI, and the SDK, on by default for Teams plans.
Key Takeaways
- Auto mode is now driven by a dedicated classifier, not simple heuristics, trained on 600,000+ real coding requests and validated across millions more in production.
- Three selectable modes (Intelligence, Balance, Cost) let users trade off quality and spend explicitly instead of accepting one default Auto behavior.
- Auto Balance reportedly beats Opus 4.8 on user satisfaction while costing about 37% less per commit, per Cursor's own published numbers.
- Evaluation relied on live A/B testing, tracking user satisfaction and code-retention rates rather than static offline benchmarks, including accounting for cache-invalidation costs when switching models mid-session.
- The rollout is broad, not desktop-only, covering desktop, web, iOS, the CLI, and the SDK simultaneously at launch.
- Enterprise admins retain override power, including mode restrictions and model allow/block lists, though Grok 4.5 can never be fully excluded, guaranteeing a low-cost tier always exists.
Sources & Mentions
2 external resources covering this update
Auto Mode Gets a Real Brain
Cursor's Auto mode has always promised to pick the right model for the job, but until now it leaned on simpler heuristics. With this release, Cursor replaced that logic with Cursor Router, a purpose-built classifier trained on over 600,000 live coding requests and validated across millions more in production. The router reads four signals for every request, the query itself, the surrounding context, the complexity of the task, and its domain, and combines that with what Cursor has learned about how each underlying model actually performs on similar work. Simple edits get routed to price-efficient models, UI and styling work goes to models known for stronger design taste, and complex, long-horizon engineering problems are sent to frontier reasoning models.
Three Optimization Modes
Rather than a single one-size-fits-all Auto setting, Cursor now exposes three distinct modes along a cost-to-intelligence spectrum:
- Intelligence targets frontier-level output quality comparable to the most capable premium models, aimed at the hardest tasks.
- Balance matches the models most developers already reach for on a daily basis, while trimming cost.
- Cost prioritizes token efficiency while still aiming for solid output quality.
According to Cursor, Auto Balance mode now scores above Opus 4.8 on user satisfaction while costing meaningfully less per commit ($4.63 vs. $7.34 in Cursor's own comparison), and Auto Intelligence lands close to Cursor's top-tier Fable model at roughly 60% lower cost for teams.
How It Was Validated
Instead of relying purely on offline benchmarks, Cursor evaluated the router through online A/B testing across millions of live requests, tracking real signals like user satisfaction and how often generated code was actually kept rather than reverted. That approach also accounts for practical costs simpler routing systems tend to ignore, such as the cache-invalidation penalty of switching models mid-conversation. Early access testing with dozens of enterprise customers reportedly showed 30-50% lower costs versus premium-model-only usage, with comparable or better code retention.
Rollout and Controls
Cursor Router is live now across desktop, web, iOS, the CLI, and the SDK. It's enabled by default for Teams plans, and Enterprise admins can turn it on from the dashboard, with the ability to set per-team defaults, restrict which of the three modes are available, and allow-list or block specific underlying models from being selected. Grok 4.5 cannot be excluded from routing, ensuring a budget-friendly tier is always available regardless of admin configuration.