Best LLM Router / Method
How we compare LLM routers
Most “best LLM router” lists rank tools without saying why. Here the criteria come first, every option is held to the same ones, and every claim links to the public page it came from with the date it was checked.
The five criteria
- 01
Model coverage
Which models you can reach through it, and from which providers. We record what the vendor's docs say and link the live model list, because catalogs change weekly. A count is only quoted when the vendor publishes one, and then as the vendor's figure.
- 02
Latency overhead
The time the layer adds on top of a direct provider call. Few vendors publish it. Where one does, we quote it as theirs; where none does, we say so and tell you how to measure it yourself.
- 03
Pricing model
How you pay, not a price list: priced per model, provider price plus a fee on credits, no markup on tokens, or open source where you pay providers and hosting. Fees on your own keys are noted. Rates are linked, never copied, because they change.
- 04
Privacy terms
Whether prompts and responses are stored, for which product, and whether zero data retention covers inference. We read the page for the exact product, since terms often cover some APIs and not others.
- 05
Fit for your workload
Where each option fits best and where it fits worse, written from the four criteria above and the features its docs describe (fallbacks, web search, self-hosting). Every option gets a weaker-fit line.
Router, gateway or model host
A router decides where each request goes (which model, which provider) and usually sells you the inference through one account. A gateway sits in front of provider accounts you already hold and adds control: keys per team, budgets, logs, caching, rate limits and fallbacks, while each provider still bills you. A model host runs the weights and sells tokens.
Every entry is labelled as a hosted router, a hosted gateway, a self-hosted gateway or an open-source gateway, so you compare like with like. Model hosts are not ranked. Definitions for these and related terms are in the glossary.
How claims are verified
- Every fact comes from the vendor’s public docs, pricing page, licence or repository, and links to that page.
- Every fact carries the date it was last checked against that page.
- A vendor’s own performance figure is labelled as the vendor’s, never presented as a measurement made here.
- Where a fact isn’t published, the cell says “not verified” or “not published” and names where to check. Nothing is estimated.
- No star ratings, scores or reviews. The verdicts say which option suits which need, and why.