BiOS Adaptive

BiOS Adaptive is one endpoint that stands in for the whole catalog. Send a request with model: "bios-adaptive" and the platform picks the right frontier model for it — quality, speed, and budget in balance. There are no routing settings to manage, and your code never changes as the catalog improves.

It speaks the same OpenAI-compatible API as every other model on Run BiOS, so adopting it is a one-word change.

Calling BiOS Adaptive

curl https://api.runbios.ai/v1/chat/completions \
  -H "Authorization: Bearer bios-your_serverless_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bios-adaptive",
    "messages": [{"role": "user", "content": "Summarize this quarter's revenue trend in two sentences."}],
    "max_tokens": 256
  }'

Everything else — authentication, streaming, the usage block, error shapes — is identical to a named model. See Serverless Inference API.

Quality Levels

Adaptive takes an optional per-request quality dial, reasoning_effort:

{
  "model": "bios-adaptive",
  "reasoning_effort": "high",
  "messages": [{"role": "user", "content": "Prove or disprove: this function is O(n log n) for all inputs."}]
}

The dial is per request, so one key can serve both a latency-sensitive chat surface (low) and a nightly analysis job (high).

Pricing

BiOS Adaptive has no single per-token rate: the price follows the model each request lands on, which is why the pricing page shows a range rather than a number. You are debited per token from your wallet exactly like a named model, and every response carries the real token counts in usage. Prompt caching is billed separately at the published cache-read rate.

Adaptive vs Named Models

• Choose BiOS Adaptive when you want strong answers without model-ops: mixed workloads, user-facing apps, or anything where "best available for the request" beats "the same model every time".
• Choose a named model when you need a pinned identity: reproducible evals, a model-specific capability, a fixed price point, or a compliance requirement to name the model in use.