FAQ

A marketplace for AI inference. Independent providers serve models, and you call all of them through one OpenAI-compatible API at https://api.numes.ai/api/v1, paying from a USD or Bitcoin balance. See Quickstart and Principles.

Request and response shapes, errors and routing controls follow OpenRouter, so most OpenRouter code works after changing the base URL and key. Unknown fields are rejected instead of ignored, each key is priced in USD or BTC, and several OpenRouter features don't exist. Compatibility lists every difference.

Any client that speaks OpenAI Chat Completions, including the official OpenAI SDKs. OpenCode, Pi and OMP start preconfigured with numes run. See Integrations.

No. The catalogue shows models, endpoints, prices and measured performance without signing in. You need an account and a key to call the API.

GET /models lists only models priced in your key's currency, so USD and BTC keys see different lists. See Which models appear.

When no endpoint can serve a request, including an unknown or misspelled model, inference returns 503 provider_unavailable. Copy IDs from GET /models, and check that the model supports the features you send, such as tools.

Yes. Take the endpoint's tag from GET /models/{author}/{slug}/endpoints and send provider: { "order": ["<tag>"], "allow_fallbacks": false }. See Provider routing.

No. A request names exactly one model, and fallback happens only between that model's endpoints, before output starts. Implement model fallback in your client; see Model fallback.

Every offer is reviewed and verified before it takes traffic, and the platform runs additional unannounced checks. Providers that misrepresent a model are suspended and lose their earnings. See Verification and trust.

Add USD credits by card, or fund a BTC balance over Lightning or on-chain, in Billing. A platform fee is added when you buy credits, and the purchase dialog shows it before you pay. Requests are charged at the providers' listed rates. See Credits and payments.

A request that fails without delivering output isn't charged. Otherwise you're charged for output delivered to you, including part of an answer delivered before an error. GET /generation is the authoritative record; see Usage accounting.

Usually one of these:

• Currency mismatch. Your credits are in USD but the key is in BTC, or the reverse.

• The hold is larger than your available balance. Each request holds an upper bound of its cost before it runs, and running requests hold funds too. A lower max_tokens doesn't shrink the hold for the same offer.

• A spending limit is used up on the key or the account.

See 402: insufficient credits.

Unused card credits can be refunded if you contact Support within 24 hours of the purchase. Bitcoin payments and platform fees aren't refundable. See the Refund Policy.

Send numeric sampling hints such as frequency_penalty as JSON numbers. Endpoints vary in which hints they accept; use provider.require_parameters: true when a parameter is required. See Parameters.

It reached its output limit: your max_tokens, or the endpoint's own maximum when you omit it. Raise the limit up to the model's top_provider.max_completion_tokens.

No. It's a correlation ID, and reusing one that's already recorded returns 400. Send a fresh ID on every attempt or omit the header. See Request and generation IDs.

Not from a third-party website: other origins are rejected with 403 cors_forbidden. Call the API from your server, and never ship a key to a browser.

No. The platform keeps metadata such as model, provider, timing, usage and cost, but no content. See Data retention.

The provider that serves a request receives its content. Each endpoint declares whether it collects or retains data and where it processes it, and you can route on that with data_collection, zdr and regions. See Provider data policies.

No. Create a new key in the other currency.

Yes. Give the key a spending limit that resets daily, weekly or monthly, or never. Owners and administrators can also set an account-wide limit. See API keys.

Open Support in the dashboard. Include the X-Request-Id of an affected request; never share your key or prompt.

Install the provider agent on Linux or a Mac with Apple silicon, then run its guided setup to choose a source, select models and set prices. The agent connects outbound over HTTPS, so you open no ports. See Become a provider.

Earnings become available after a holding period, and you withdraw them in US dollars to a bank account or in Bitcoin. See Earnings and payouts.