Free AI API
Stop paying per token for AI.
No credit card — not now, not after a quota runs out. A flat cost per request that does not move with prompt length, and credits you can earn by contributing compute rather than buy. Below: what it really costs, and the three things it does worse than a hosted API.
Why does my AI bill keep growing?
Because almost every provider bills per million input and output tokens. That sounds fair until you build something real. Retrieval stuffs documents into the prompt. Agent loops re-send the whole conversation on every step. A coding assistant pastes an entire file. None of that changes what you asked for — but all of it changes what you pay.
The result is a bill that scales with your prompt engineering rather than your usage, and which is genuinely hard to forecast. It is why teams end up rationing context, truncating history, and picking weaker models for jobs that deserved better ones.
Tenseuron does not meter tokens. The cost of a task comes from the network's configured budget, so a 200-token prompt and a 20,000-token prompt cost exactly the same.
What does one request cost?
| Per-token providers | Tenseuron | |
|---|---|---|
| Billed on | Input + output tokens | Per accepted task |
| Long prompt costs | More | The same |
| Payment method | Card, after the free tier | Wallet — no card, ever |
| Free allowance | Granted quota, then it stops | 3 tasks per network, plus credits you can earn |
Concretely: a task consumes 0.02 of the network's token. At recent token prices that is around $0.0000138 — roughly 72,000 tasks per dollar. Token prices move, so treat the dollar figure as illustrative; the 0.02 is the mechanism.
Every wallet also gets its first three tasks free on every network, enforced from real task history rather than browser state — so clearing cookies does not reset it, and you can evaluate a network before locking anything.
How do I use an AI API without paying at all?
Earn the credits instead of buying them. This is the part that has no equivalent elsewhere, and it works because of one detail: miners are paid in the same token that funds usage.
Run the CLI as a miner → serve tasks with your own model → get paid in the network token → lock what you earned → call the API with it. No money enters the loop at any point.
If you run the model locally through Ollama or LM Studio, there is no API bill on the mining side either — your inputs are hardware and electricity you already own. If you prefer a hosted provider, you supply your own key and it stays on your machine.
To be clear about what this is not: it is not passive income. The point is not that you get paid — it is that your AI bill goes to zero. If you want to compare earnings against renting a GPU on a rental marketplace, this is the wrong tool.
If my GPU can serve tasks, why would I need the API?
A fair objection, and the answer is that the two are not the same resource. Your card runs one request at a time. The network runs many in parallel. Mining converts capacity you have in the shape of idle hours into capacity you can spend in the shape of a burst — serve tasks overnight while the machine is doing nothing, then push ten thousand classifications through in the morning and have them finish in minutes rather than days.
Two smaller versions of the same point. Your machine sleeps, gets rebooted, and spends hours running something other than inference; the network does not, so work you dispatch to it continues while your card is unavailable. And the model you can host is not necessarily the model the job wants — a 7B that runs comfortably on your hardware can earn against tasks it handles well, and that credit can buy you results from miners running something far heavier than you could load.
Does my existing OpenAI code work?
Yes. Point your base URL at Tenseuron and existing client libraries work unchanged — there is an OpenAI-compatible surface alongside the native endpoint:
POST /v1/chat/completions GET /v1/models POST /api/compute # native endpoint GET /api/compute/:taskId # poll for the result
Authenticate with a bearer API key, or with a wallet signature that stays valid for five minutes. Results come back by polling, or you can supply a webhook and have them pushed to you.
Where it falls short
Three real trade-offs. If any of them is a dealbreaker for what you are building, use a hosted API — that is the honest answer:
- You cannot pin a model. You submit a task describing the job; miners choose what runs it and validators score the results. Consensus filters weak answers, but output style varies more than it would against a single provider. Buyer and creator model selection is planned, not shipped.
- No latency guarantee. A task waits for a miner to claim it. Fine for batch work, screening, and anything on a timer. Not fine for a streaming chat cursor.
- A wallet is required. No credit card is a real advantage, but it is not the same as frictionless — you need a Solana wallet, and using a network means acquiring and locking that network's token.
- The network is early. Fewer miners means less redundancy than an established provider offers. That improves as networks attract supply, but it is the current state and you should plan for it.
How to start without paying
- 1
Install the CLI
One command sets up mining: npx @tenseuronp/cli setup. No stake is required to start.
- 2
Point it at a model
Run a model locally through Ollama or LM Studio for no API bill at all, or supply your own hosted provider key — it stays on your machine.
- 3
Serve tasks
The miner polls for open work, runs it through your model, and submits the output for validation.
- 4
Spend what you earned
Accepted work pays out in the network's token. That is the same token that funds usage, so lock it and call the API without spending anything.
npx @tenseuronp/cli setup
Frequently asked questions
Is there a genuinely free AI API?
Most 'free' AI APIs give you a granted quota that runs out — a trial, some credits, or a daily request cap. Tenseuron works differently: you can earn credits by contributing compute instead of being granted them. Run the CLI as a miner, serve tasks, and you are paid in the same token that funds usage on that network. Every wallet also gets three free tasks per network to try it before mining anything.
Do I need a credit card?
No, and not later either. There is no billing step at any point — you connect a Solana wallet rather than entering payment details. That is a real difference from free tiers that ask for a card once you pass the quota, but it is not the same as costless: you need a wallet, and using a network means acquiring and locking that network's token.
What does one request actually cost?
A flat 0.02 of the network's token per task, and it does not change with prompt length. At recent token prices that works out around $0.0000138 per task — roughly 72,000 tasks per dollar. The cost comes from the network's configured budget rather than being metered per token in and out.
Why doesn't the price go up with a longer prompt?
Because it is not billed per token. Every other provider charges per million input and output tokens, so a long-context request costs more than a short one and your bill moves with prompt size. Here the per-task cost is derived from the network budget, so a 200-token prompt and a 20,000-token prompt cost the same.
If my GPU can serve tasks, why would I need the API at all?
Because a single card and a network are not the same resource. Your GPU runs one request at a time; the network runs many in parallel. Mining converts capacity you have as idle hours into capacity you can spend as a burst — serve tasks overnight, then push ten thousand jobs through in the morning and have them finish in minutes instead of days. The network also stays up when your machine sleeps or reboots, and credit earned with a 7B you can comfortably host can buy results from miners running models far heavier than your hardware would load.
Can I run it on my own hardware?
Yes. Miners can run models locally through Ollama or LM Studio, or via any OpenAI-compatible endpoint. In that setup there is no API key and no API bill at all — the only inputs are hardware and electricity you already have. If you prefer a hosted provider you supply your own key, and it never leaves your machine.
Does my existing OpenAI code work?
Yes. Alongside the native endpoint there is an OpenAI-compatible surface at /v1/chat/completions and /v1/models, so existing client libraries work by pointing the base URL at Tenseuron.
What are the catches?
Three, stated plainly. You cannot pin a model — you submit a task and miners choose what runs it, with validators scoring the result. There is no latency guarantee, because a task waits for a miner to claim it. And the network is early, so there is less redundancy than a hosted provider offers.
Try it on three free tasks
Every wallet gets three per network, no card and nothing to lock. If it does what you need, start mining and stop paying for inference entirely.