Inference network
An OpenAI-compatible API that records what it served
Use Basis to query models through an OpenAI-compatible API while the network records verifiable inference receipts and routes planned $BASIS rewards to the contributor GPUs that served the work.
Inference API contract live · runtime + token not live
The OpenAI-compatible API contract is defined; the worker backend and $BASIS token are pending. $BASIS is the planned credit and reward unit — no contract is deployed and no Bankr launch is open. Nothing here is an offer to buy any asset, and no earnings are guaranteed.
What it is
Basis is a Base-native inference network. A request enters through an OpenAI-compatible gateway, is routed to a GPU that can serve the model, and is metered token-by-token on the server. Every served job produces a deterministic receipt: an auditable accounting record of what was computed, by which worker, at what token count.
The API surface mirrors the conventions developers already use (/v1/chat/completions, /v1/models), so existing clients point at it with a base-URL change. What is different is underneath: accounting is deterministic, receipts are hash-stable, and reward routing is contributor-aware.
How credits work
Usage is planned to be denominated in $BASIS base units and priced deterministically per token — integer base units, floor division, no floating point anywhere in the money path. Prompt and output tokens are weighted separately, and a model multiplier scales the charge for heavier models.
This is planned, not live. Pricing constants are accounting placeholders until launch; the credit vault that would hold deposits is pending. No deposits are accepted until the vault is configured.
How jobs are routed
A request hits the gateway, which hands it to an orchestrator that selects a contributor GPU worker advertising the requested model. The worker streams tokens back through the gateway to the caller.
Server-counted tokens are authoritative. The network counts output tokens as it relays them, rather than trusting a worker's self-report — that count is what the receipt and any reward are computed from. Jobs served directly by the gateway (no contributor worker) are metered but earn no $BASIS reward: they are billed work, not contributed compute.
Why receipts matter
Every served job produces an inference receipt: a SHA-256 hash over the canonical JSON of the job's immutable economic core — model, token counts, multipliers, charged and reward amounts. The hash excludes mutable envelope fields (status, signature, settlement pointers) so it stays stable across the receipt's lifecycle.
Because the hash is deterministic, anyone can recompute it from the same inputs and confirm the accounting was not altered after the fact. Receipts are the audit trail that makes both billing and reward routing checkable rather than asserted. See recent receipts.
An API example
A streaming chat completion. The Authorization bearer key shown here is a placeholder — API keys are pending until the runtime and credit vault are configured.
curl https://basis.watch/api/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-basis-..." \
-d '{
"model": "basis-default",
"messages": [{ "role": "user", "content": "Summarize this transaction." }],
"stream": true
}'List the advertised models and their runtime availability:
curl -s https://basis.watch/api/v1/models
Advertised models
Every model is planned until a worker actually serves it at runtime. The multiplier scales pricing and rewards relative to 1x.
| Model id | Context window | Multiplier | Status |
|---|---|---|---|
| basis-default | 8,192 | 1x | planned |
| basis-small | 8,192 | 0.5x | planned |
| basis-large | 32,768 | 3x | planned |
Current launch status
- Inference runtimeruntime pending — API contract live, backend absentPending
- $BASIS tokenaddress — pending deploymentPending
- Credit vault contractPending
- Reward distributor contractPending
$BASIS is planned for launch on Bankr. Bankr launch — pending. See launch preparation.
For how GPU workers join and earn, see /workers. For settlement batches on Base, see /settlement. Pricing constants are accounting placeholders, not final pricing; $BASIS is planned, not live.