Verify what served your request

Discounted API access is only worth using if the model behind it is the one you asked for. This page lists what AnyPath guarantees, what it changes, and the checks you can run yourself.

What reaches the provider

• Official endpoints only. A route can only point at the provider’s own endpoint for its access type. Suppliers cannot enter a custom upstream address.
• The model you name. The model in your request maps to one exact upstream model on the route that serves it. AnyPath does not substitute a cheaper or smaller model.
• Your request body. AnyPath does not add, remove, or rewrite prompt content, and does not inject system messages or output limits. The one exception is a documented protocol conversion: for example, a Chat Completions request served by an Anthropic route is translated to the Messages format. Fields that cannot be translated faithfully are rejected instead of dropped.

What comes back

• The provider’s own response. Response bodies and streaming events are passed through. When your request used a model alias, the model field is set back to the name you asked for.
• Provider usage numbers. The token counts you are charged for are the ones the provider reported in that response. If a response has no usable usage figures, the request is not charged.
• Headers. Provider response headers are forwarded, except those that identify the supplier’s own account or describe its quota. X-Request-Id is AnyPath’s ID for your request.

What is recorded for each request

Open a request in Console → Requests to see:

• the provider that served it and the exact upstream model
• input, output, cache, and reasoning token counts as reported by the provider
• the official list price and the multiplier applied
• what you were charged and how the request settled

Checks you can run

• Count tokens independently. Compare the usage in a response with the provider’s own tokenizer or its token-counting endpoint for the same input.
• Use provider-specific features. Prompt caching fields, reasoning or thinking blocks, tool calling, and structured output behave as they do against the provider directly.
• Run your own evaluation set. Send the prompts you already use to judge a model and compare the answers with a direct call.
• Watch the public numbers. Live supply and price shows success rate and first-byte latency per provider.

What AnyPath does not claim

• Suppliers are independent account holders. AnyPath verifies that a credential works against the official endpoint and serves the listed models; it does not control how a supplier configured their own provider account, including logging options on that account.
• Aggregator providers such as OpenRouter and Vercel AI Gateway are offered only for models pinned to a named upstream. Their own routing sits between AnyPath and the model.
• Consecutive requests are not guaranteed to reach the same supplier account, so a provider-side prompt cache may not always hit.