- Who it helps
- Researchers and developers evaluating model-routing systems.
- Availability
- November 3 paper is public. A deployable routing service was not verified.
The comparison
The authors collected 374,401 scored runs across 452 configurations and 25 benchmarks spanning several modalities and tasks.
An upper bound is different from a working router
The reported oracle improvement is 5.4 percentage points, while learned routing gains 0.3 points across 16 benchmarks. An oracle can choose the best result after seeing outcomes; a deployed router must choose beforehand.
What you can do
- Compare the learned-router result and inference cost before using the oracle figure in a deployment estimate. Open reference ↗
Where the evidence stops
Original abstract checked November 4. No routing code or benchmark was executed, and the numbers are author-reported.
Sources read. Primary sources read November 4, 2026 UTC. No installation, model inference or complete workflow executed.
Sources & further reading
Paper · paper
RouteAnything compares models and reasoning policies ↗
Identifies the subject, documented capability or specific verification limit.
Pengfei Zhou and coauthors · page checked 2026-11-04
Coverage note: Independent attention and end-to-end execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
RouteAnything evaluates how much a system can gain by choosing a model and reasoning policy for each task.