Cursor has made Cursor Router usually obtainable for Groups and Enterprise plans. The system is a classifier that inspects every request earlier than a mannequin runs, then dispatches it to the mannequin finest suited to that particular activity. The cursor staff stories frontier-quality efficiency at 60% financial savings in on-line A/B exams, and 30–50% financial savings for early-access enterprise accounts.
The issue it targets is a spend sample slightly than a functionality hole. Cursor states that roughly 60% of its builders choose a single mannequin as a day by day driver. Routine work due to this fact will get accomplished at frontier costs, and AI spend grows quicker than output high quality. Routing is Cursor’s reply to that mismatch.
What the classifier really reads
Cursor Router will not be a fallback chain or a retry mechanism. It’s a classifier skilled on 600k+ reside requests, evaluated in a web-based A/B check throughout hundreds of thousands of reside requests, and optimized for consumer satisfaction (AFC) as its reward sign.
For every request, the router analyzes 4 inputs: question, context, activity complexity, and area. It combines these with discovered data of every mannequin’s conduct. Cursor publishes three routing guidelines that observe from that classification:
- Easy work goes to probably the most price-efficient fashions.
- UI updates go to the mannequin with one of the best style.
- Advanced, long-horizon issues go to frontier reasoning fashions.
That third rule carries the burden of the fee argument. The financial savings don’t come from downgrading laborious issues. They arrive from eradicating routine work from frontier pricing whereas the tough tier stays intact.
One implementation element deserves consideration from anybody who has constructed a router. Cursor Router is cache-aware in each coaching and analysis. It’s skilled on a dataset the place routing produces cache misses, and the reported value financial savings embody the price of these cache misses. Switching fashions mid-conversation invalidates immediate cache, and that value is actual. Routers that ignore it overstate their financial savings.
The classifier was additionally designed for mannequin churn. Cursor states it may well replace the router as newer fashions ship, which issues in a market the place the frontier strikes month-to-month.
Interactive Explainer
Inside Cursor Router: classification, tiers and printed outcomes
Each determine under is taken from Cursor’s launch publish and changelog (July 22, 2026). Values Cursor didn’t publish are marked as such.
Decide a activity — see how the classifier reads it
Step 1 · Alerts learn earlier than any mannequin runs
▼
Step 2 · Routing tier chosen
Cursor publishes three routing guidelines and doesn’t identify a particular mannequin per activity kind. The tier proven is Cursor’s personal printed class; the pattern duties are illustrative functions of these three guidelines. Cursor states the classifier reads question, context, activity complexity and area, mixed with what it is aware of about every mannequin’s conduct.
Auto mode — three factors on the fee–intelligence Pareto frontier
Printed value per commit (early entry measurement)
Price-per-commit figures are Cursor’s printed early-access observations. The projection is straight multiplication of these unit figures — it’s arithmetic, not a Cursor forecast, and assumes a continuing activity combine. Cursor stories GPT-5.6 Sol matched the price of Intelligence mode however with decrease consumer satisfaction; it didn’t publish an actual determine for it.
| Ingredient | What Cursor stories |
|---|---|
| Coaching knowledge | 600k+ reside requests |
| Reward sign | Consumer satisfaction (AFC) |
| Analysis | On-line A/B check throughout hundreds of thousands of reside requests, not offline evals |
| High quality metric 1 | Consumer satisfaction — agent success categorized from consumer responses. Shifting to the following function is a powerful constructive; correcting the agent is a powerful unfavourable. |
| High quality metric 2 | Hold price — how a lot agent-generated code stays within the codebase over time |
| Metric monitor file | Each metrics used to judge each mannequin launch and harness change for the previous 9 months |
| Cache dealing with | Cache-aware in coaching and analysis. Skilled on a dataset the place routing causes cache misses; reported financial savings embody cache-miss value. |
| Platform scale | Cursor routes a whole lot of hundreds of thousands of coding requests per week throughout fashions and suppliers |
| Downside being solved | ~60% of Cursor builders choose one mannequin as a day by day driver, placing routine work on frontier pricing |
Cursor’s acknowledged motive for selecting on-line A/B exams: offline evals are restricted by small dimension, distance from actual utilization, and the problem of lowering success to a rubric — and so they omit the additional cache-miss value incurred when switching fashions mid-conversation.
Why on-line A/B exams as a substitute of offline evals
Cursor intentionally averted offline evals as its major measurement. Its acknowledged reasoning is that offline evals undergo from small pattern dimension, distance from real-world utilization, and the problem of lowering success to a rubric. Additionally they omit the cache-miss value incurred when switching fashions.
Actual routing occurs throughout a dialog, not a single flip. Builders write code, ask follow-ups, hit errors, and proceed, typically throughout a whole lot of requests in every week. The router should resolve each which mannequin to select and when to modify.
Two high quality metrics carry the analysis:
- Consumer satisfaction: agent success categorized from consumer responses. Shifting on to the following function is a powerful constructive sign. Correcting the agent is a powerful unfavourable one.
- Hold price: how a lot agent-generated code stays within the codebase over time.
Cursor staff states it has used each metrics to judge each mannequin launch and harness enchancment for the previous 9 months. It is a significant credibility marker: the metrics predate the product they’re now getting used to justify.
Three modes, and the numbers behind them
Auto mode now exposes three optimization settings that transfer the consumer alongside the fee–intelligence Pareto frontier.
- Auto Intelligence lands close to Fable on consumer satisfaction at about 60% decrease value for groups. In opposition to Opus 4.8, it lifts satisfaction about 15% at almost the identical value.
- Auto Steadiness lands above Opus 4.8 on consumer satisfaction at about 36% decrease value. In opposition to GPT-5.6 Sol, it delivers comparable satisfaction at a decrease spend price.
- Price mode is described nearly as good high quality reaching the very best obtainable intelligence whereas optimizing token spend. Cursor printed no A/B high quality or value figures for it.
Price per request is barely a part of the image, so Cursor additionally measured value per commit:
| Mannequin / mode | Price per commit |
|---|---|
| Auto Steadiness | $4.63 |
| Auto Intelligence | $6.76 |
| Opus 4.8 | $7.34 |
| Fable 5 | $12.69 |
GPT-5.6 Sol matched the price of Intelligence mode however produced decrease consumer satisfaction. Cursor didn’t publish an actual per-commit determine for it.
