Cursor Router: choose Cost, Balance, or Intelligence with an eval gate
Quick answer
Cursor launched Cursor Router on July 22, 2026 for Teams and Enterprise. Auto mode now classifies each request by its query, context, task complexity, and domain, then sends it to an underlying model. You choose one of three optimization modes:
- Cost for bounded, repeatable work where a lower accepted-task cost matters most.
- Balance as the starting default for mixed product work.
- Intelligence for difficult diagnosis, architecture, and long-horizon implementation where failure or rework is expensive.
Do not enable a mode organization-wide because of a headline savings percentage. Cursor's early-access results are its own production measurements, not a guarantee for your repositories. Build a fixed task set, reveal the routed model during the canary, compare cost per accepted change, and keep a fixed-model fallback.
Who this guide is for
This guide is for engineering leads and small teams using Cursor Agents who need to control AI coding quality and spend without asking every developer to pick a model for every prompt.
If you are still choosing fixed models, use the same fixtures from the Gemini 3.6 Flash routing guide and Grok 4.5 coding-agent eval. Model routing does not replace the untrusted-repository sandbox gate.
What changed
Cursor Router is a routing layer, not a new foundation model. Cursor says the classifier was trained on more than 600,000 live requests and evaluated across millions of requests. It can switch the underlying model per request and is cache-aware in its training and production evaluation.
| Mode | Start here when | Main risk | Required evidence |
|---|---|---|---|
| Cost | Tests, mechanical edits, documentation, and bounded cleanup | Cheap attempts create more corrections | Accepted-change rate and rework time |
| Balance | A team has a mix of routine and moderately complex work | An average hides weak task categories | Results split by workload class |
| Intelligence | Repository diagnosis, migrations, security review, and long tasks | Frontier-quality routing exceeds the budget | Cost per accepted task and failure severity |
Cursor reports that early-access customers saved roughly 30% to 50% against routing all traffic to Opus 4.8, and that online tests found larger savings in selected comparisons. Treat those figures as vendor evidence. Reproduce the decision on your own accepted work.
Balance and Intelligence are billed at the routed model's rate. Cursor also gives admins mode restrictions, model allow/block lists, defaults, and soft or hard enforcement. The routed model is hidden by default but can be displayed. Teams have Router enabled by default; Enterprise admins can enable it. Availability covers desktop, web, iOS, CLI, and the SDK.
Choose a mode by the cost of being wrong
Use task consequence before task length:
| Workload | First mode | Escalate when | Keep fixed-model control |
|---|---|---|---|
| Rename, formatting, test generation | Cost | The first correction exceeds the expected savings | Fast low-cost model |
| Small feature with clear tests | Balance | The agent misses acceptance criteria twice | Current daily-driver model |
| Intermittent bug or multi-file migration | Intelligence | Start here; downgrade only after repeated success | Best proven reasoning model |
| Security-sensitive change | Intelligence | Never relax permissions to rescue a route | Reviewed fixed model plus human approval |
| Long conversation with valuable context | Current model | Start a new routed session instead of switching mid-chat | Existing session and branch |
The last row matters because caches are model-specific. Cursor's harness team recommends staying with one model through a conversation unless there is a reason to switch. A fresh subagent or new session is easier to compare and roll back.
Build a 30-task routing eval
Create three buckets of ten real, sanitized tasks:
- Routine: tests, copy changes, renames, and one-file fixes.
- Product: small features, UI changes, API wiring, and refactors.
- Hard: ambiguous bugs, migrations, security boundaries, and multi-file work.
Run the current fixed model and each allowed Router mode from a fresh branch and fresh conversation. Keep repository state, instructions, permissions, and verification commands identical. Record this evidence without prompt contents or secrets:
task_id: product-07
route: auto-balance
routed_model: visible-during-canary
accepted_without_correction: true
kept_diff_after_24h: true
tool_errors: 0
verification: passed
latency_seconds: 84
cost_usd: 0.42
rollback_needed: false
Calculate total route cost / accepted tasks, not average request cost. Also track correction turns, kept diff after a day, verification pass rate, tool errors, p95 latency, and rollback frequency. A cheap route that creates rejected patches is not cheap.
Six-case canary gate
| Case | Probe | Pass condition |
|---|---|---|
| Routing fit | Run all 30 tasks with the routed model visible | Each task class has an explicit default and escalation path |
| Quality | Blind-review diffs and run deterministic tests | Acceptance and verification do not regress beyond the agreed threshold |
| Continuity | Compare fresh sessions with one controlled multi-turn task | Cache misses and model switches do not create hidden latency or context loss |
| Permissions | Repeat one denied network or write action | The route cannot widen tools, secrets, or approval boundaries |
| Cost | Compare total spend, corrections, and accepted changes | Cost per accepted task fits the mode budget |
| Rollback | Disable Router for a canary group | New sessions return to the fixed model and open work remains recoverable |
Promote one group at a time: 5%, 25%, then the target share. Keep the routed model visible during the canary. Stop on permission drift, test regressions, repeated correction loops, unknown model changes, or a budget breach.
Common mistakes
- Treating Cursor's aggregate savings as a repository-specific forecast.
- Sending every task to Intelligence because it sounds safest.
- Measuring token spend without measuring accepted changes and rework.
- Hiding the routed model before debugging and cost attribution are stable.
- Blocking an underlying model without testing how that changes Router behavior.
- Comparing modes in different conversations, branches, or permission profiles.
- Switching a valuable long-running conversation mid-task instead of starting a fresh routed session.
FAQ
Is Cursor Router available on individual plans?
The launch announcement says Teams and Enterprise. Do not assume individual Pro availability from the presence of Auto in another plan.
Which mode should a small team start with?
Start with Balance for mixed work, then move bounded tasks to Cost and high-consequence tasks to Intelligence only when your eval supports it.
Does Router remove the need to choose models?
It removes repeated per-prompt selection, but admins still choose allowed models, modes, defaults, visibility, and enforcement. You still need a fallback and an audit trail.
Top comments (0)