Cursor Releases Cursor Router: An Application-Level Classifier Delivers Frontier Code Quality at 30–50% Lower Cost
Cursor has made Cursor Router widely available for Teams and Enterprise applications. The system is a system that evaluates each request before the model runs, and then sends it to the model best suited for that particular task. The cursor team reports quality performance bordering on 60% savings in online A/B testing, and 30–50% savings for early access business accounts.
The problem being addressed is the spending pattern rather than the energy gap. The cursor says almost that 60% of its drivers choose one model as a daily driver. So normal work is being completed at marginal rates, and the use of AI is growing faster than the quality of the output. The route is the cursor response to that mismatch.
What the producer reads
Cursor Router is not a reverse chain or retry method. It is the classification system that you are trained by 600k+ live requeststested in online A/B testing on millions of live applications, and optimized for user satisfaction (AFC) as a reward signal.
For each request, the router evaluates four inputs: question, context, task complexity, and background. It combines this with the learned behavior of each model. Cursor publishes the following three routing rules from that partition:
- Easy operation goes to the more expensive models.
- UI updates go to the model with the best taste.
- Complex, long-horizon problems go to frontier thinking models.
That third rule weighs on the cost argument. Savings do not come from reducing difficult problems. They arise from removing the normal function from the marginal values while the critical phase remains intact.
One implementation detail needs attention of anyone building a router. Cursor Router is cache-aware for both training and testing. It is trained on a dataset where the route produces a cache that is not available, and the cost savings reported include the cost of those temporary caches. Changing models in mid-conversation invalidates the data warehouse, and those costs are real. Ignored routers forward their money.
The classifier is also designed for the churn model. Cursor says it can update the router as it ships new models, which is important in a market where the border moves every month.
Interactive Descriptor
Inside Cursor Router: classification, categories and published results
All values below are taken from Cursor’s post and changelog (July 22, 2026). Values that the Cursor did not print are marked as such.
Choose an activity — see how the classifier learns it
Step 1 · Signals are read before any model starts
▼
Step 2 · Route category is selected
The cursor publishes three routing rules and specifies a specific model for each type of operation. The section shown is the section published by the Cursor; sample jobs are examples of those three rules. The cursor says that the separator is reading question, context, task complexity and backgroundcombined with what it knows about the behavior of each model.
Default mode — three points on the cost–intelligence Pareto frontier
Published cost per commitment (early access rate)
Cost per commit calculations are a visual representation of Cursor early access. Projection is the exact multiplication of those unit calculations – it’s arithmetic, not Cursor prediction, and it takes a constant combination of functions. Ccursor reports GPT-5.6 Sol it matched the cost of Intelligence mode but with lower user satisfaction; it did not publish its exact figure.
| A feature | Reported by Cursor |
|---|---|
| Training data | 600k+ live requests |
| Reward signal | User Satisfaction (AFC) |
| Testing | Online A/B testing on millions of live applications, not offline tests |
| Quality metric 1 | User satisfaction — an agent’s success is divided by user feedback. Moving on to the next feature is a pretty solid thing; fixing agent is a strong negative. |
| Quality metric 2 | Save rate — how much code generated by an agent stays in the codebase over time |
| Metric track record | Both metrics were used to assess each model’s launch and change of direction over the past nine months |
| Warehouse management | Cache-aware in training and testing. Trained on a dataset where the route causes cache misses; reported savings include the cost of cache misses. |
| Field scale | Cursor delivers hundreds of millions of coding requests per week across all models and providers |
| The problem is solved | ~ 60% of Cursor developers choose one model as a daily driver, putting the usual work on borderline values |
Cursor’s stated reason for choosing online A/B tests: offline evals are limited by their small size, distance from real use, and the difficulty of reducing success in a rubric – and leave additional cache-miss costs incurred when changing models during the discussion.
Why online A/B testing instead of offline evals
The cursor has deliberately avoided offline levels as its main measure. Its stated hypothesis is that offline evals suffer from small sample size, distance from real-world application, and the difficulty of reducing success to a rubric. They also omit cache-miss costs incurred when switching models.
The real route happens in a conversation, not a single turn. Developers write code, query tracking, debug, and move on, often across hundreds of requests per week. The router must decide both which model to choose and when to switch.
Two quality metrics to evaluate:
- User satisfaction: the success of the agent is divided into the responses of the user. Moving on to the next feature is a strong positive signal. Fixing agent has a strong negative.
- Keep up the standard: how much agent-generated code stays in the codebase over time.
The cursor team says it used both metrics to evaluate each model’s launch and harness development over the past nine months. This is a logical credibility marker: pre-product metrics are used to justify.
Three ways, and the numbers behind them
The default mode now presents three optimization settings that move the user closer to the cost–intelligence Pareto frontier.
- Auto Intelligence lives near legend in terms of user satisfaction at about 60% lower costs for groups. Against the Opus 4.8, it raises satisfaction by about 15% at about the same cost.
- Auto balance It outperforms Opus 4.8 in user satisfaction at about 36% lower cost. Compared to the GPT-5.6 Sol, it delivers the same satisfaction at a lower spend.
- Costs The mode is described as a good quality that achieves the highest intelligence available while optimizing the use of tokens. Cursor has not published its A/B quality or cost figures.
The cost per request is only part of the picture, so the Cursor is also rated cost per obligation:
| Model / mode | Cost per obligation |
|---|---|
| Auto balance | $4.63 |
| Auto Intelligence | $6.76 |
| Opus 4.8 | $7.34 |
| Myth 5 | $12.69 |
GPT-5.6 Sol matched Intelligence mode costs but produced lower user satisfaction. Cursor did not publish the exact amount of each of its commitments.
