Questions, Answered

For chat completions, streaming, embeddings and model listing, yes: point any OpenAI SDK at our base URL with a sparehash key. Tool calling and image inputs aren't supported yet and return a clear error.

Open-weight models such as Gemma 4 and 3, Qwen 3.8, 3.6 and 3.5, gpt-oss 20B and 120B, Qwen3 Coder, DeepSeek R1, Phi-4, Mistral Small, Ministral, Granite, Nemotron, GLM, OLMo and Llama, plus embedding models. The markets page lists them with live prices, context length and license. OpenAI, Anthropic, Google, Mistral, xAI, DeepSeek and Alibaba Cloud models are on sale too, from sellers with spare capacity, at a discount off list price.

Keys set to verified-only are routed exclusively to nodes with a strong spot-check record. It costs 1.5× the normal price.

Each key has a requests-per-minute and tokens-per-minute limit. When you hit one you get a 429 with a retry-after header.

If it drops before the first token, the request moves to another node and you never notice. If it drops mid-answer, another node continues from the text you already received. You're charged only for tokens you actually receive; a request that fails completely is free.

The network sends hidden test prompts with fixed settings and compares the answers to a reference. Nodes that fail lose traffic and are eventually banned.

Credits and payouts are USDC on Base: transfers cost about a cent, settle in seconds and work in most countries. Card payments may come later.

Sign in with your wallet, then send USDC on Base from that same wallet. Credits appear after the transfer confirms. If you buy USDC on an exchange, withdraw it to your wallet on Base first.

No. Credits are prepaid service credits: they can't be withdrawn or transferred. Top up only what you plan to use.

A tiny amount of ETH on Base (a few cents) pays the network fee for your deposit transaction. Signing in is free.

You choose the limits: maximum resource use, a schedule such as idle-only or nights-only, and a pause toggle. The app stops serving as soon as you need your machine.

No. The app makes an outbound connection to the network, like a browser does.

It depends on demand, your hardware tier and your uptime. The app shows an estimate after benchmarking. Compare it with your electricity cost before leaving it on around the clock.

Usually, yes. Payouts are income for a service you provide. We collect a W-9 or W-8BEN before your first payout and issue tax forms where required. You're responsible for your own taxes.

Yes, on the Sell tab of your dashboard: add your key, pick models, set a discount off the provider's list price, and each model is tested before it goes live. Only official providers are supported. Before you do, check your agreement with the provider: standard API terms (OpenAI, Anthropic, Google and others) generally don't allow reselling or sharing access, and a provider can suspend a key or account that breaks its terms. Only list accounts where you're allowed to, such as reseller or enterprise agreements that permit it.

You get 85% of what buyers pay for tokens served with your key; sparehash keeps the rest. Your provider still bills you its normal rate for those tokens, so selling below list only makes sense if your own cost is lower: unused credits, committed spend, or volume discounts. Earnings are held 7 days, then paid in USDC.

Sellers never receive requests: sparehash calls the provider directly with the seller's key, with store turned off. Because some provider consoles can show request logs, the privacy filter also masks credentials and personal details on these requests.

If the provider rejects your key, all its offers come off the market immediately and the key is marked "Rejected". Offers that fail now and then are rested with growing pauses and marked unhealthy; use Re-test to check them again.

Not your secrets or personal details. Before a request leaves sparehash, the privacy filter replaces credentials, emails, phone numbers, card numbers and similar data with placeholders, and puts the real values back into the answer for you. The node only ever sees the placeholders. The rest of the text is visible to the model, and so, in principle, to someone who takes apart their own computer. The sparehash app sandboxes the model and never stores prompts, and nodes are tested around the clock. How the filter works.

Names, street addresses, and secrets written in plain words. For health, legal or financial records, call the provider directly from your own account rather than through a shared network.

Yes, occasionally. If a node fails mid-answer, the conversation continues on a second node, which then sees it too, with the same values masked.

Anything illegal, including child sexual abuse material, which we report to the relevant authorities. Each model's own license and acceptable-use policy also applies. Keys used for abuse are suspended.

We can't accept funds from, or pay, sanctioned addresses or people in sanctioned countries. Wallets are screened.