A slot on the fleet, $2 an hour.

What the slot is

Model

DeepSeek V4.1 Flash

Context

1,000,000 tokens in a single request

Vision

Images in, text out

Tools

Function calling and reasoning modes, as the model exposes them

Concurrency

1 — one request at a time. A second parallel request waits for the first

Retention

None. Prompt and completion content is never logged, never stored

Metered

Token counts and wall-clock seconds, in aggregate, for billing. Not content

Endpoint

POST https://api.snail3d.com/v1/chat/completions

Model name

snail-reserve-top

Auth

Authorization: Bearer sk-snail-… — the key you get on this page

Prepaid packs

Every pack is the same slot at the same rate: $2.00 an hour. Buy the hours you want; they sit on the key you claim below.

The pool is finite. How many slots exist and how many are free is read live at the top of this page.

Get your key

1. Pay on Polar.Checkout sends you straight back to this page with your checkout id in the address.
2. The key is minted.A slot comes out of the pool and an API key is created for the hours you bought. This page watches for it, checking once every two seconds.
3. Copy the key.It is shown once, here, and never again. Save it before you close the tab.
4. Start sending.Put the key in an Authorization: Bearer header against the endpoint above.

Send a request

OpenAI-compatible. Any OpenAI client works — point the base URL at https://api.snail3d.com/v1 and set the model to snail-reserve-top.

curl https://api.snail3d.com/v1/chat/completions \
  -H "Authorization: Bearer sk-snail-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "snail-reserve-top",
    "messages": [
      {"role": "user", "content": "Hello."}
    ],
    "stream": true
  }'

Streaming works the way it does anywhere else. One request at a time per slot: send a second one while the first is open and it waits its turn.

How it compares

OpenRouter lists three providers for V4.1 Flash, all metered per token. Their published rates per million tokens:

Rates as published on OpenRouter. Decart's listing shows the input rate only. Where a provider charges above those rates, that is their price, not ours.

Where the flat hour wins

Per-token providers bill every token you resend, and a long context is resent on every turn. Hold the full 1,000,000-token context and send one turn every 30 seconds and that is 120 turns in an hour, or 120 million input tokens:

Wafer

120 × $0.05 = $6.00 for the hour, before output

Sail

120 × $0.08 = $9.60 for the hour, before output

Decart

120 × $0.09 = $10.80 for the hour, before output

Here

$2.00 for the hour, however large the context and however many turns

The other direction is true too, and you should hear it: with short prompts and a small context, per-token billing is cheaper than any flat hour. The flat hour is for the workload where a large context stays in play and the meter would keep running — and, unlike those three, nothing you send here is retained or used to train anything.

Zero data retention, plainly

Your prompts and the model's completions are not written to a log and not written to a store. They pass through the server to answer your request and they are gone. There is no prompt archive, no completion history, and nothing to hand to a trainer.

What is kept is the bill: aggregated token counts and wall-clock seconds per key, and the key itself. That is what the meter runs on, and it is the whole of it. Nothing is sold, and there is none of yours to sell.