You.com Launches Inference Router: 12 Open-Weight Models With 99.9% Uptime SLA

You.com launched Inference Router, a managed, OpenAI- and Anthropic-compatible API that puts 12 open-weight models behind one endpoint. The pitch is the paperwork: a 99.9% minimum uptime SLA on pay-as-you-go, zero data retention covering both model calls and web search, 100% US hosting including failover, and no platform fee.

That matters because a lot of teams have already decided open-weight models are good enough for real work. What stalls them is the security review: where does the data live, is it stored, who can touch it. This product is aimed squarely at that gap.

We like this direction. Open-weight models from Moonshot, Alibaba, Z.ai, DeepSeek, OpenAI and others are now competitive enough that the question for most companies is operational, not technical. A router that answers the compliance questionnaire up front is the kind of boring infrastructure that actually moves AI into production.

Be clear-eyed about what is being sold, though. By MarkTechPost's own comparison, per-token prices sit near market. Kimi K3 is listed at $2.85 in and $14.25 out per million tokens, only about 5% below Together AI. The value is the SLA, the retention policy and the US-only footprint. That is a fair thing to charge for, but buyers should price it as insurance, not as a bargain.

The benchmark claims deserve the usual asterisk. You.com says GLM-5.3 with its search tied GPT-6.1 Sol's 86.53% on a 1,329-question LiveNewsBench run at a lower cost per question, and that DeepSeek V4.1 Flash came within about two points of Claude Sonnet 5.5 at roughly 77% less per question. Those numbers are self-reported, on a news-retrieval task where You.com's own search product is part of the setup. They say something real about cheap models plus good retrieval. They don't say open models have caught frontier models across the board.

The longer trajectory is the one I'm watching. If routers like this become the default way enterprises consume open weights, the model makers fade into interchangeable suppliers and the leverage moves to whoever owns the failover, the search layer and the compliance promise. Failover here runs to partners including Parasail, Wafer, Alibaba Cloud, Meta and Inception, so a zero-retention guarantee is only as strong as every partner honoring it. That is a trust chain worth auditing, not assuming.

Developers typically test open-weight models on aggregators like OpenRouter, then hit a wall when security reviewers ask about data processing agreements and hosting location. Dedicated providers can clear those bars but usually want large commitments. You.com routes requests to its own capacity first and fails over to vetted US-hosted partners at the same published price, and offers per-key spend caps on SOC 2 compliant infrastructure.