Sign InOpen Brain
CursorEngineering PostOfficial Source

How Cursor Router chooses the right model for the task

Cursor Router learns task complexity and model fit from production behavior, showing why agent routing should include correction signals, cache costs, and per-task performance.

Cursor · Aug 6, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Cursor routes each turn in two stages: Compass estimates complexity, then a production-derived taxonomy selects among eligible frontier models. **Auto Intelligence costs 68% less than Fable**, while **Auto Balance costs 41% less than Opus 4.8** and reports higher satisfaction.

Practical Implication

Builders implementing model routing should learn from real task outcomes, including corrections and task progression, rather than rely only on benchmarks. Include token use, cache misses, model-switching costs, confidence thresholds, and an explicit per-turn budget.

Agent-Ready Context
Cursor routes each turn in two stages: Compass estimates complexity, then a production-derived taxonomy selects among eligible frontier models. **Auto Intelligence costs 68% less than Fable**, while **Auto Balance costs 41% less than Opus 4.8** and reports higher satisfaction.

Builders implementing model routing should learn from real task outcomes, including corrections and task progression, rather than rely only on benchmarks. Include token use, cache misses, model-switching costs, confidence thresholds, and an explicit per-turn budget.

Cursor's satisfaction measure is inferred from subsequent user behavior, so it is a proxy rather than a direct quality score. The reported results come from Cursor traffic and may not transfer to another workload, user population, or pricing mix.
Connected Context · Feed7 Judgment

This turns model routing into a per-turn, budget-aware learning problem grounded in production behavior, not a static benchmark lookup. It contrasts directly with task-consistent routing: Cursor accepts switching costs to match each turn, while acknowledging that inferred satisfaction and workload-specific economics limit how broadly its reported savings can be applied.

TRACE-ROUTER: Task-Consistent and Adaptive Online Routing for Agentic AITRACE-Router keeps one backend for an entire task, directly contrasting with Cursor’s per-turn selection and making model-switching cost versus finer-grained adaptation the central design choice.TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated ReasoningTurnSight reinforces the value of turn-level signals, but uses execution hindsight for training while Cursor infers routing quality from subsequent user behavior.Data and Environment Curation for Post-Training LLMs — Mahesh Sathiamoorthy, Bespoke LabsBoth argue that improvement should come from real task outcomes and carefully designed data signals rather than benchmark performance alone; the post-training work also underscores the need to check whether learned routing transfers across distributions.
Context Map
agentcodingdata#model-selection#harness-engineering#agent-reliability
Uncertainty
Cursor's satisfaction measure is inferred from subsequent user behavior, so it is a proxy rather than a direct quality score. The reported results come from Cursor traffic and may not transfer to another workload, user population, or pricing mix.