Arbiter 27B ยท API
One endpoint, the same model Orrery runs on.
The Arbiter API is an OpenAI-compatible chat completions endpoint. Any client that speaks the OpenAI format works unchanged: point it at https://api.ephemerent.com/v1, use arbiter-flash-27b as the model, and authenticate with a key from your account page.
Chat completions at /chat/completions, streaming with stream: true, tool calls in the OpenAI schema. The model list at /models returns one entry.
Arbiter 27B: a recurrent-depth Qwen3.8-27B that runs three looped passes at prefill and decodes in a single pass. Read how it is built and measured.
Each request costs prompt tokens times the three loop passes plus completion tokens, priced at provider cost. One credit is one millionth of a dollar. There are no billed reasoning tokens: Arbiter reasons in latent space.
API keys draw first from the same monthly Arbiter 27B pool as the desktop: 14M on Pro, 35M on Max, 70M on Ultra. When a month's pool runs out, requests fall through to credits you have bought, which never expire and work across both models. Nothing overages: with neither balance the API answers 429 rather than billing you. Organizations share one pool across seats.
Your first request
curl https://api.ephemerent.com/v1/chat/completions \
-H "Authorization: Bearer $ARBITER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "arbiter-flash-27b",
"messages": [{"role": "user", "content": "Find the rule: 2, 6, 12, 20, 30"}],
"max_tokens": 400,
"chat_template_kwargs": {"enable_thinking": false}
}'Python and JavaScript clients from OpenAI work as they are: set base_url to https://api.ephemerent.com/v1 and api_key to your key. Keep enable_thinking off; the loop replaces visible chain-of-thought.
Usage, keys and limits
GET /v1/usage with your key returns this month's compute units, the loop pass count and the retention mode. The same balance shows on your account page next to the desktop pool.
Create and revoke keys on the account page. A key shows its prefix after creation; the full key is shown once. Revoking takes effect on the next request.
Credit packs are a one-time purchase on your account page: 10M for $20, 25M for $45, 60M for $100. They sit in a rollover balance the monthly pool is always spent before, so buying early never wastes them.
Per-key requests and tokens per minute scale with the plan, and every key is capped by the monthly pool. Over the limit the API answers 429 with a budget_exhausted error; nothing is billed beyond the pool.
Data
API requests follow the same rule as the desktop: Arbiter learns from sessions by default under an anonymous session id, and Zero data retention is a switch on your account. With it on, the service keeps counts and an access-log line, never prompts or outputs. Details on the privacy page.