Evidence at a glance
Moving the Bottleneck from Code to Trust
You.com has introduced Inference Router, a managed API for production use. Developers can call 12 open-weight models through OpenAI- and Anthropic-compatible interfaces, and can enable web search in a request. The service is aimed less at the question of how to call a model than at what happens after a team has tested open models and needs to pass security review and operate reliably. It promises at least 99.9% uptime, zero data retention, US-based hosting, pay-as-you-go pricing, and no platform fee.
Those promises target costs that are easy to underestimate during a migration. Teams can compare models quickly on an aggregator, but production review raises different questions: whether prompts are retained, who processes requests, where traffic goes during an outage, and whether usage commitments are required. Inference Router bundles these operational arrangements with the model endpoint. Its value therefore depends on the full deployment boundary, not just the model list or API compatibility.
One Endpoint, Multiple Inference Paths
Requests first go to You.com’s own Inference Direct capacity. If that capacity is unavailable or full, the service fails over to vetted partners. The launch material names Parasail, Wafer, Alibaba Cloud, Meta, and Inception, with Prime Intellect described as coming soon. You.com says these routes are US-hosted and subject to its zero-retention rules, and that failover does not change the price for a given model.
This design removes capacity management and provider switching from the caller’s code, but it does not eliminate the underlying supply chain. A unified API and fixed per-model prices simplify integration and budgeting. Platform owners still need to know whether performance is consistent across routes, what triggers failover, and how provider changes are communicated. The available material does not disclose capacity allocation, failover latency, or service metrics for individual partners. The existence of failover alone does not mean every request receives identical availability guarantees.
Zero Retention Defines a Boundary, Not an Absence of Records
You.com describes zero data retention as covering both model calls and web searches: prompts and outputs are processed in memory and not stored. Prompt hashes, de-identified text, and disk-backed prompt caches are explicitly excluded. The service still retains operational metadata such as token counts, timestamps, and account IDs. Accounts run on SOC 2-compliant infrastructure, and the material says a data processing agreement is available on request.
ZDR should therefore be read as a commitment about handling specified content data, not as a claim that the system leaves no records or that the entire data path meets every organization’s requirements. A review should clarify how the contract defines processing and storage, which fields appear in troubleshooting logs, and whether partners are bound by equivalent terms. The material describes the policy’s scope but does not detail audit methods, retention exceptions, or customer-configurable access controls. Those points still need contractual and technical review.
Search Changes More Than the Call Pattern
Developers can declare a web_search or web_extract tool so the router runs searches for the model, then returns a web_results object with citations. The documentation says each search can return up to 20 results and that the default is 30 search rounds. Search and model inference run on the same server. You.com also reports up to 15% lower latency in internal tests, without providing test conditions or independent validation in the supplied material.
Combining retrieval and inference behind one endpoint reduces application orchestration and gives open-weight models a ready-to-call search path. The tradeoff is that outputs now depend on search results, the number of search rounds, and citation handling, so evaluations cannot compare models alone. Teams should include web retrieval in their data-flow review. Although the ZDR policy is said to cover search requests, the handling of request content, returned citations, and operational metadata should still be mapped to the application’s data classifications.
Read the Price Advantage Alongside the Evaluation
Published prices show that differences between models can be larger than differences between platforms. Per million tokens, GLM 5.3 costs $1.35 for input and $4.25 for output, while DeepSeek V4.1 Flash costs $0.30 and $1.20. Context windows range from 125K for GPT-OSS 120B to 1M for several models. Kimi K3, at $2.85/$14.25, is only about 5% below the $3.00/$15.00 listed for Together AI. A unified endpoint and no platform fee therefore do not mean that token costs are always substantially lower.
You.com reports a different cost measure from its LiveNewsBench run of 1,329 questions. In its self-reported test, GLM-5.3 with You.com search scored 86.53% at $0.0485 per question, compared with $0.0920 for Claude Sonnet 5.5. This is a useful lead about a combined model-and-search setup, not a general result established by the supplied evidence: You.com reported the test, no independent replication details are provided, and the comparison is between an open model with search and a closed model with built-in search. Technical leads should rerun evaluations on their own task sets, measuring quality, latency, and full request cost before migrating.