Which API should you route your model calls through?
All four put many models behind one key. They differ less than any of them advertise — and where they do differ, it is not always in our favour.
- Enso600+
- OpenRouter500+(80+ providers)
- Portkey1,600+
- LiteLLM1,800+(140+ providers)
Side by side
Opens on the 1 row we ran ourselves. Reported rows — what each vendor says about itself — are one click away.
| Enso vs OpenRouter, Portkey, LiteLLM | Hanzo Enso | OpenRouter | Portkey | LiteLLM | Provenance |
|---|---|---|---|---|---|
| GPQA-Diamond, run by us | The router beside single models from other labs on the same 198 questions, every row a research run with its accuracy and what its tokens cost at list prices[ 2026-10-15 ] | not measured by us | not measured by us | not measured by us | we ran this |
| Enso vs OpenRouter, Portkey, LiteLLM | Hanzo Enso | OpenRouter | Portkey | LiteLLM | Provenance |
|---|---|---|---|---|---|
| GPQA-Diamond, run by us | The router beside single models from other labs on the same 198 questions, every row a research run with its accuracy and what its tokens cost at list prices[ 2026-10-15 ] | not measured by us | not measured by us | not measured by us | we ran this |
| Models reachable | 600+ | 500+ (80+ providers) | 1,600+ | 1,800+ (140+ providers) | they published this |
| OpenAI-compatible endpoint | Yes — /v1/chat/completions | Yes | Yes | Yes | they published this |
| Native Anthropic /v1/messages | Yes — same key | Yes — Anthropic Messages format, tools and extended thinking | Yes — /v1/messages routes to any of their models | Yes — serves the format at /v1/messages | they published this |
| Cost known before the call runs | Yes — model calls are quoted from above before they happen, and an agent that would breach its cap is refused rather than discovered on an invoice | No — blocks once a limit is reached (403), Enterprise plan only. No pre-flight quote published | No — budget limit expires the key once reached, Enterprise plan. No pre-flight quote published | No — caps budgets and stops at the cap. No pre-flight quote published | they published this |
| Routing quality published by the vendor | Yes — GPQA-Diamond, 198 graduate-level questions, every row a research run | No — their Benchmarks API scores the models, not the router | Not published | Yes — Auto Routing Cost Ladders: 58.2% savings, 46.8% win rate vs flagship | they published this |
| Runs on your own infrastructure | Yes — the whole cloud is one binary and one API, on your own Kubernetes or your laptop | No — a hosted service | Self-hosting offered at enterprise tier | Yes — open source, self-hosted free forever | they published this |
GPQA-Diamond, run by us: One set of 198 graduate-level questions through the router and the field, read from the research API. We have not run the other three routers on this set, so their cells say that rather than borrowing a marketing number to sit beside a measurement.
Reading this honestly
This page is not arranged to make Hanzo look best.
Where OpenRouter, Portkey, LiteLLM are ahead, that is on the page. The bet is that a comparison you can check is worth more than one that only says good things about itself.
A row marked we ran this is a number produced by our own harness, and the link goes to the run. A row marked they published this is each vendor’s own claim, taken from the page linked and dated the day it was read — we relay those, we did not verify them, and they are their numbers and not ours. Where a vendor has not published something the cell says not published rather than being left blank, because a blank reads as zero.