You.com Launches Managed Inference Router for 12 Open-Weight Models

Inference Router lists both OpenAI-compatible and Anthropic-compatible routes at https://api.you.com. You.com said existing SDK code can move to the service by changing the base URL and API key. The service is pay-as-you-go with no platform fee, and each model has one published per-token price. You.com said the router targets a gap between aggregators and dedicated providers: teams test open-weight models on aggregators, then face security reviews over zero data retention, U.S. hosting, access controls and data processing agreements, while dedicated providers that clear those requirements usually require large commitments. The service is fully managed, with no GPUs to manage.

Accounts run on SOC 2-compliant infrastructure, with a data processing agreement available on request, and customers can set spend caps per API key and view daily usage by model, cost and provider. The 12 live models are all U.S.-hosted, according to the report. They range from GPT-OSS 120B with a 125,000-token context window to Kimi K3 and Qwen3.8 Max with 1 million-token contexts. Kimi K3 is described as a 2.8 trillion-parameter model, and Qwen3.8 Max as a 2.4 trillion-parameter mixture-of-experts model. Prices per 1 million tokens are $2.85 input and $14.25 output for Kimi K3; $1.85 and $5.55 for Qwen3.8 Max; $1.35 and $4.25 for GLM 5.3; $1.65 and $3.65 for DeepSeek V4 Pro; $0.30 and $1.20 for DeepSeek V4.1 Flash; $0.10 and $0.35 for Mercury 3; and $0.097 and $0.73 for GPT-OSS 120B. Muse Wonder, Muse Spark 1.3, GLM 5.3 Flash, Qwen3.8 Flash and MiniMax M3 complete the list.

Requests first go to You.com Inference Direct, the company's own inference capacity, and fail over to vetted partners if that capacity is down or full. The official post names Parasail, Wafer, Alibaba Cloud, Meta and Inception as partners, with Prime Intellect coming soon. All partners are 100% U.S.-hosted and must meet zero-data-retention rules, and failover does not change the bill, according to the report.

The zero-data-retention policy covers model calls and web search requests. Prompts and outputs are processed in memory and not stored; You.com keeps only operational metadata such as token counts, timestamps and account IDs. The policy excludes prompt hashes, de-identified text and disk-backed prompt caches.

For web search, developers declare a web_search or web_extract tool, and the router runs searches for the model. The documentation shows up to 20 results per search and a default of 30 search rounds. Responses return a web_results object with citations. Search runs on the same server as the model, and You.com said internal tests show up to 15% lower latency.

You.com also published results from its LiveNewsBench test, saying open-weight models plus its search matched frontier models at lower cost. GLM-5.3 with You.com search scored 86.53% at $0.0485 per question, compared with $0.0920 for Claude Sonnet 5.5 in self-reported results. On per-token pricing, the report said prices sit near market levels; Kimi K3 at $2.85 input and $14.25 output is 5% below Together AI's $3.00 and $15.00.

Editor's Summary

You.com launched Inference Router, a managed API for 12 open-weight models with U.S. hosting, zero data retention and a 99.9% uptime SLA. The service targets teams whose open-weight pilots have stalled in security review and includes built-in web search, published per-token pricing and partner failover. You.com's LiveNewsBench results claim frontier-level performance at lower cost for some models.