Run a workload on Ayni.

POST /v1/workloads returns a price and an ETA built from the live supply — how many eligible devices are online and what they actually sustain. Accepting it runs the batch and charges you once, at the quoted price.

curl https://api.ayni-ai.com/v1/workloads \
  -H "Authorization: Bearer $AYNI_KEY" -H "Content-Type: application/json" \
  -d '{"model":"qwen2.5-0.5b-instruct-q4_k_m",
       "items":[{"messages":[{"role":"user","content":"Summarize: ..."}],"max_tokens":160},
                {"messages":[{"role":"user","content":"Summarize: ..."}],"max_tokens":160}]}'
# -> { "id":"wl_...", "estimate":{"eta_seconds":...,"eligible_nodes":3},
#      "price":{"total_usd":0.42,"breakdown_usd":{...}}, "accept_url":"/v1/workloads/wl_.../accept" }

curl -X POST https://api.ayni-ai.com/v1/workloads/wl_.../accept \
  -H "Authorization: Bearer $AYNI_KEY"
# -> { "object":"workload.result", "items":[ ...in submission order... ], "charged_usd":0.42 }

Prefer to place the batch yourself? POST /v1/batch takes the same items, fans them out immediately, and meters pay-as-you-go.

Not in a hurry? Add "spot": true. Spot workloads run on whatever capacity on-demand traffic leaves free — priced at 60% of on-demand, with a wider completion-time band, and Ayni may interrupt and requeue them.

Point any OpenAI SDK at the coordinator and pass your key. Streaming and blocking /v1/chat/completions and /v1/models work as you’d expect.

curl https://api.ayni-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $AYNI_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Provider-Trust-Level: device_attested" \
  -d '{"model":"qwen2.5-0.5b-instruct-q4_k_m",
       "messages":[{"role":"user","content":"Hello"}]}'

Prepaid credits, billed per million tokens by model class, the same unit every hosted model uses. Chat and batch are metered at the rate card; a quoted workload adds coordination, a failure buffer and payment cost on top of the provider’s share, which lands about 20% above it. Launch rates are placeholders and will be set with the community before general availability. Per 1M tokens, input / output:

• Micro (≤3B, live today): $0.02 / $0.08 · spot 60% of that
• Small (3–8B): $0.05 / $0.20
• Medium (12–30B quantized): $0.15 / $0.60
• device_attested ×1.4 · confidential ×3

Compare on that number, not on a per-item price. A 220-token log line with a one-word answer is about 226 tokens, so one million of them cost about $6 on the micro class: $0.027 per million tokens. For reference, hosted open models of 1–8B list at roughly $0.02–0.10 per million tokens, the commercial nano tier at $0.10–0.40, and mid-tier cloud models at $0.30–2.50 (list prices, approximate, batch tiers about half).

70% of what you pay goes to the provider that served you; the rest funds the coordinator, the community treasury, and the AI-stakeholder allocation.

The live model is 0.5B. It is fast and cheap, and it does not reason. The workloads that fit are mechanical, short, constrained and large:

• Normalize and extract. Raw audit, syslog or CloudTrail events into one JSON schema: actor, action, target, source, time. Severity by deterministic rules on top.
• Closed-set labelling. Tickets, alerts or documents against a fixed list of ten to twenty categories with worked examples.
• Embeddings for semantic search (on the roadmap: one endpoint).
• Integrity by redundancy. The same item on three devices in three countries with a majority vote, in one call.

Judgment (“is this an incident?”) belongs to the small and medium classes on laptops and GPUs. Every price view returns the price per million tokens, and a built-in self-test scores the model on labelled lines before any money is spent.