You.com’s Inference Router Sells the Paperwork, Not the Weights

Author

AI News Editorial

Published

2026-10-24 10:15

You.com has launched Inference Router, a managed API that puts 12 open-weight models behind a single OpenAI- and Anthropic-compatible endpoint. The models are not the interesting part. The contract is.

The service ships with a 99.9% minimum uptime SLA backed by service credits, zero data retention across both model calls and web search requests, 100% US-hosted infrastructure including failover, a DPA on request, and no platform fee — pay-as-you-go, 0% markup, against OpenRouter’s 5.5%.

Why an API for models you can self-host

You.com frames the problem as stalled migrations. Teams prototype open-weight models on an aggregator, then hit the security review: reviewers ask about ZDR, US hosting, access controls and DPAs. Answers vary by provider and sometimes by model, so the evaluation dies in the queue.

Routing itself is straightforward. Every request goes first to Inference Direct, You.com’s own capacity; if that is down or saturated, requests fail over automatically to a vetted partner. The model lineup runs from GPT-OSS 120B with 125K context up to Kimi K3 and Qwen3.8 Max with 1M context, and web search runs on the same servers as the model rather than through a third party.

The benchmark claim to read carefully

On You.com’s LiveNewsBench run, the vendor reports that GLM-5.3 paired with You.com search scored 86.53% at $0.0485 per question, against $0.0920 for Claude Sonnet 5.5 — roughly half the cost for a higher score. Those are self-reported numbers on a self-run benchmark.

The pricing spread tells a different story at the model level: Kimi K3 is listed at $2.85/$14.25 per million tokens, only about 5% below Together AI’s $3.00/$15.00. At the token level, open-weight hosting has largely converged on price. What has not converged is the legal and operational wrapper around it.

That is the real product here. Serving an open-weight model is a commodity; the enterprise complaint is never that the GPU is too slow, it is that nobody will sign the DPA. You.com is selling the signature.

Related: Introducing You.com Inference Router