How InferKey Works

How your requests are relayed, why you get genuine models, and what that means for billing

How your request flows

your app → InferKey → official model capacity → Claude / GPT

InferKey verifies your API key, forwards your request over official model channels, and streams the response back byte-for-byte. We do not rewrite your prompt, inject instructions of our own, swap the model, or post-process the content. If capacity for a model hiccups, requests fail over automatically.

Your prompts and completions are not stored by InferKey — we record only per-request metering (model, token counts, cost, duration), which is exactly what you see on your Usage page.

Genuine models, verifiable

You get the official models — same weights, same quality, same token accounting. Ways to convince yourself:

Run the same prompt against the official API and InferKey side-by-side and compare outputs and reported token usage.
Streaming responses carry the provider's native event format (e.g. Anthropic's message_start / content_block_delta events) untouched.
Per-request token counts on your Usage page match what the SDK reports in the response.

The Claude channel: system prompt & caching

Claude traffic travels over the official Claude Code CLI channel. This keeps the supply stable and compliant — and it has two practical consequences you should know about.

Your request	What is added	Overhead
Full Claude Code client (system + tools)	Nothing — forwarded as-is	~0
Your own system prompt / tools	Identity core only; yours kept unchanged	~140 chars
No system prompt, no tools	Identity core only	~140 chars

Billing: the channel content carries a 1-hour prompt cache that is managed for you. The first request in a conversation pays a one-time cache write; requests within the hour read it at ~10% of the base input rate; in long conversations it amortizes to near-zero. You will see this as cache read/write tokens in your usage data.

Capability difference: forced structured output is not available on the Claude channel (a channel capability, not a different model). Tool calling, streaming, vision, and extended thinking all work.

Billing model

Prepaid credits, metered per token at the published rates. Your per-token rate is locked when you purchase. When your balance runs out, requests stop with a 402 — you can never overspend. Every request appears on your Usage page with its exact cost.

Questions? Contact us.