How to read an inference receipt
September 16, 2026 · 4 min read
Every Arctic call ends with a receipt. Here is what each number means, and which ones matter when you choose a model.
A receipt has six fields: input tokens, output tokens, time to first token, total time, who served the call, and what the same call would cost on a paid API. On Arctic the call itself is free, so the numbers that matter are the waiting and the quality.
Tokens in and out
Input tokens are your prompt plus any instructions; output tokens are the answer, including hidden reasoning when the model thinks before it writes. Two models can give the same answer with very different output counts, and more tokens usually means more waiting.
First token versus total time
Time to first token is how long you stare at an empty box. Total time is when the answer is complete. For chat, first token matters most; for batch jobs, only the total counts. A reasoning model often has a slow first token and a reasonable total.
Served by
Free tiers get busy. When one host says no, Arctic sends the call to another host of the same model. The receipt names the host that actually answered, so you can tell a slow model from a busy host.
Paid APIs
The last field multiplies your tokens by the model's list price on paid APIs. It is not a bill. It shows what the same work would cost you elsewhere.
What a receipt does not tell you
- Whether the answer is right. Read it.
- Whether the model would do as well on your next prompt. Run a few.
- Whether the speed will hold at busy hours. Check the Ledger's average route time over a week.
A good habit: pick three prompts that look like your real work, run them through Crossing with several models, and keep the receipts. The fastest model that gets all three right is usually the one to put behind your key.