One operator, exact models, reserved by the hour.

What a reservation fixes

Pricing

Priced by the dedicated GPU your models need, per compliant hour, post-paid. Non-compliant hours cost nothing. Machine-readable at GET /api/v1/reservations/pricing.

The setup fee is $5 + one hour at the tier price, once per booking (GPU warm-up). The operator, the models and the SLA threshold are agreed per reservation before it is booked. Write to hello@infernetprotocol.com.

Keys for your customers

Each reservation can have any number of API keys (prefix ifr_res_). You issue and revoke them yourself with the reservation's buyer token (see below), one per customer if you like; the shared limits still cap all of them together. A key is shown once. Revocation takes effect on the next request. Keys work with any OpenAI client:

export OPENAI_BASE_URL=https://infernetprotocol.com/v1
export OPENAI_API_KEY=ifr_res_...

curl -sS $OPENAI_BASE_URL/models -H "Authorization: Bearer $OPENAI_API_KEY"   # only the reserved models

curl -sS $OPENAI_BASE_URL/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"<reserved model id>","messages":[{"role":"user","content":"Classify: ..."}]}'

When a limit is hit

Refusals are HTTP 429 with headers that say which limit and when to retry:

Successful responses carry X-RateLimit-Remaining-Requests, X-Infernet-Shared-Remaining-Requests and the concurrency limits. Every refusal is counted in the report as a limit hit.

How availability is proven

Every minute of the window, whether or not anyone sends a request, the control plane probes the pinned target and stores one record for that minute:

• Operator endpoint: GET /models must list every reserved model id. With completion probes on, a one-token completion on each model must also succeed and report that model.
• Infernet node: the node's signed heartbeat must be under 90 seconds old, report itself available, and list every reserved model as served.

A minute is confirmed when its probe succeeded and no real request failed upstream in that minute. A minute with no record (for example if the prober itself did not run) is unconfirmed; it is never assumed to be fine. An hour is compliant when its confirmed minutes reach required_minutes_per_hour. Hours with no requests are judged the same way, so an idle reserved hour that stayed up is compliant.

The report you can check

Each reservation comes with a buyer token (prefix ifr_buy_) scoped to that one reservation. With it you can issue and revoke keys, and read the per-hour verdicts, the failures behind them, every per-minute record, and usage (requests, tokens, limit hits). It cannot change the reservation's terms or see anything else:

curl -sS https://infernetprotocol.com/api/v1/reservations/<id>/report?minutes=1 \
  -H "X-Buyer-Token: ifr_buy_..."

curl -sS https://infernetprotocol.com/api/v1/reservations/<id>/invoice \
  -H "X-Buyer-Token: ifr_buy_..."

Or from the CLI:

INFERNET_BUYER_TOKEN=ifr_buy_... infernet reservation key-issue <id> --label customer-a
INFERNET_BUYER_TOKEN=ifr_buy_... infernet reservation report <id>
INFERNET_BUYER_TOKEN=ifr_buy_... infernet reservation invoice <id>
INFERNET_BUYER_TOKEN=ifr_buy_... infernet reservation watch <id>    # live view

Post-paid billing

Nothing is charged up front and nothing is charged by this system. The invoice view lists each reserved hour with its status and computes amount_due = compliant hours × price_per_hour + setup_fee (the setup fee only once an hour is compliant). Non-compliant and cancelled hours are owed nothing. The figure is final only once every hour has ended. You verify it against the report, then pay as agreed.

API, CLI and MCP

The same operations are in the CLI as infernet reservation … (run infernet reservation --help) and as MCP tools via infernet mcp, e.g. claude mcp add infernet -- infernet mcp.

Back to the documentation.