We document the sources, observations, and judgment behind each review so readers can decide what applies to their own systems.
ModelRouter.com is maintained by industry professionals with an interest in AI systems optimization. Our reviews and guides reflect our personal observations and opinions based on the information available at the time of publication. They do not represent the views of our employers, clients, or other professional relationships.
We aim to be thoughtful, measured, and consistent, but your results may vary. Different workloads, constraints, risk tolerance, geography, and organizational priorities can reasonably lead to a different conclusion. Conduct your own research before implementing a recommendation, including an assessment of operational risks and vulnerabilities.
Provider reviews begin with a detailed assessment of the product's functionality and the documentation published by the service provider. We examine the product's stated capabilities, architecture, deployment options, pricing, model and provider coverage, routing controls, security posture, operating model, and intended use cases.
Where a review includes direct product observation, a setup flow, or a measured result, that activity is identified in the review. Documentation-based findings are not presented as measured performance, and the methodology does not imply that every product has been exercised through the same code-level test.
A review explicitly identifies a setup flow, interface behavior, request, or measurement observed by the author. This label applies only to the activity described.
A primary provider source supports the statement. This establishes what the provider says is offered, not how well it will work for every customer.
A claim is supported by another primary source, an observable result, or a repeatable public record.
The author's conclusion based on the facts, documentation, observations, and tradeoffs disclosed in the review.
The capability was outside the review's available evidence. No claim is made about its quality or performance.
The information needed to assess the item was not public or could not be confirmed at the time of review.
Each category is scored from 1.0 to 5.0 after its listed components are assessed against the available evidence. The overall score is the unweighted mean of all seven category scores and is rounded to one decimal only after calculation. A score is withheld when the evidence is insufficient.
Decimal scores are editorial distinctions within these bands, not laboratory measurements. Use the category evidence and product fit, not a tenth of a point, as the deciding factor.
Data handling, privacy controls, compliance posture, access controls, and security safeguards.
Breadth of supported models, providers, endpoints, and deployment options.
Fallbacks, load balancing, policy controls, and intelligent model-selection capabilities.
Latency, uptime, resilience, failure handling, and available operational evidence.
Setup, documentation, APIs, debugging tools, observability, and day-to-day usability.
Governance, support, deployment flexibility, auditability, and organizational controls.
Pricing transparency, platform fees, cost controls, and the value delivered for the spend.
The provider operating a hosted router can have access to inbound and outbound messages passing between an application and a model provider. That can include prompts, responses, request metadata, tool definitions, tool calls, and other sensitive material. A compromised or untrustworthy intermediary can create confidentiality and integrity risks.
Recent research documents malicious intermediary attacks involving payload manipulation and secret exfiltration, and explains why normal transport encryption does not prove that the response received by a client matches the upstream model's response. Review the paper on malicious LLM router attacks and our security and privacy guide before selecting a provider for sensitive workloads.
Many products in this category are early stage and change frequently. High-change fields such as pricing, model access, policies, and product capabilities are checked on a rolling basis, but a provider may change after a review is published. Confirm current details and perform an assessment for your own circumstances before adopting a service.
Material corrections are noted on the affected page. If you have source-backed evidence, a compelling observation, or a well-supported alternative view, we would like to hear from you.