Provider keys and pricing

Provider keys, any OpenAI-compatible provider, and pricing per call, per 1,000 tokens or per job unit.

Provider keys

The AI page holds one key per provider, per environment. Paste it, make a test call, and it is set. A key is 8 to 512 printable characters. Replace it by pasting again. A job provider's key is pasted like a chat key; its test call is one authenticated read that spends nothing.

The key is write-only. No response, no audit record and no page shows more than its last four characters. The AI page shows the last four characters of a key of 16 or more characters (a shorter key shows dots) and the last call. A provider that echoes the key in a refusal reaches the app as ***<hint>.

In Live, removing a key asks for a code from your email first, and your app's AI features stop the moment it is removed.

Optionally, list up to 20 models the app may call; the test call uses the first. Any other model answers model_not_allowed. The model name is the provider's, 1 to 80 characters of letters, digits, dots, colons, underscores and hyphens. Google needs one; OpenAI and Anthropic refuse a missing one themselves.

Any OpenAI-compatible provider

Chat runs on a standard, not a list. Besides OpenAI, Anthropic and Google, a chat tool, an alias or a raw call runs on any server that speaks the OpenAI Chat Completions format, as provider openai-compatible. On the AI page, add a provider and pick Any OpenAI-compatible provider, then fill in four things: a name (what you call it), the base URL from the provider’s docs, the API key, and the models you allow (optional; a tool names its model either way). One per environment.

Gemmein sends POST <base URL>/chat/completions with Authorization: Bearer <key> and the standard body (model, messages, max_tokens from the tool’s output cap, stream) and nothing else, passes the answer back in the standard’s shape, and meters from its usage field. It checks the base URL on every call: https only, a host name and never an IP, never a gemmein.com host, every address it resolves to public, and the connection pinned to the address it checked. It never follows a redirect. The key is encrypted and write-only, and a new base URL is saved only with its key pasted again.

{
  "label": "Quick answer",
  "provider": "openai-compatible",
  "model": "meta-llama/llama-3.3-70b-instruct",
  "credits": 1,
  "promptTemplate": "{{question}}",
  "inputs": [{ "name": "question", "type": "text", "required": true }]
}

A tool names one of the endpoint’s own model ids (up to 120 characters; slashes, :, @ and + allowed), or an alias to one. It reads text only until it ticks image under accepts; tick it only for a model that reads images. It takes no audio, video or document inputs.

Pricing

A tool holds one credit price. For a price that depends on the plan, make two tools, each with its own access requirement. How credits are bought, granted and spent is on Credits.

Per call

call is the default unit. The tool's credits are spent before the request is forwarded; the default tool spends one. A person short of the price is answered 402 credits_exhausted before anything is sent, with the balance in the message.

Per 1,000 tokens

With unit: "tokens", per is 1,000 by default. Before the call the tool reserves a ceiling: the price of the request's bytes ÷ 3 plus the output cap (4,096 when the tool sets none). When the answer ends cleanly, the call settles at the provider's usage, up to the ceiling, and what was not used returns. Per-token pricing is for text: a tool that takes files is priced per call.

The call costs the ceiling when the provider reports no usage, when a stream does not end cleanly, or when the call is abandoned mid-answer. A call abandoned before the provider answered costs the input estimate.

Per job unit

A generate tool meters images, seconds or characters (or call: one price per run). A transcribe tool meters seconds (or call). A tool's unit is the one its model meters: an image model per image, a video or transcription model per second, a speech model per character, or call. Any other unit is refused when the tool is saved, since a unit nothing meters would charge every person the ceiling.

A metered job tool reserves credits × ceil(bounds.maxUnits / per) when its run starts, and settles at what the provider metered when the run succeeds.

Refunds

Credits spent before the request are refunded only when the provider fails before its first byte, with a non-2xx answer or no answer at all. The response then carries x-gemmein-credit: refunded. A stream that dies after the first byte is not refunded, and hanging up early does not refund.

Response headers

Every answer that passed the spend carries x-gemmein-credits-remaining. x-gemmein-tool names the tool; absent on the implicit default.

Usage & billing

The Usage & billing page counts AI calls for the last 30 days: spent from your people's credits, priced by your provider. Gemmein meters the calls; your provider bills the tokens.