BiOS Adaptive — enterprise grade inference

One endpoint. Every model. The right one for each request.

Adaptive routes each prompt across the open-model catalog to the model that best fits it. You send one API call — you pay only for the model that answered.

The problem

Most prompts are billed at rates they never needed

01

Everyone defaults to the strongest model for everything

It is the safe choice — and it means most prompts are billed at frontier rates they never needed.

02

One model cannot be right for every prompt

A classification call and a multi-step reasoning call have nothing in common. Paying frontier rates for both wastes money; paying budget rates for both wastes quality.

03

Providers go down

Hard-coding one model id into your product is a single point of failure you only discover during the outage.

How it works

One call in, the best-fit answer out

Your request arrives

The same OpenAI-compatible schema you already use. No SDK migration, no new client.

Adaptive picks the model

It predicts the best-fit model for that specific input and routes the request to it — frontier strength when the prompt demands it, budget speed when it does not.

The answer returns

Through the same endpoint you called. If the chosen model is unavailable, the request reroutes instead of failing.

1M token contextTool callingOptional reasoning

Who it is for

The same call, worth something different to each seat

For engineers

One integration, no model babysitting. Point your existing OpenAI client at the endpoint, set the model id to bios-adaptive, and change nothing else.

For product teams

A quality floor on hard prompts without paying frontier prices on easy ones. Simple prompts land on budget models; hard prompts land on strong ones — automatically.

For the business

You pay the rate of the model that actually served each request — the band below, with no routing fee stacked on top of it.

Pricing

The band is the savings mechanism, made visible

BiOS AdaptiveRate, per 1M tokens
Context window1M tokens
Input$0.14$1.25
Cached input$0.14
Output$0.28$3.95

The rate varies because the model varies — and you only ever pay for the model that answered. A prompt routed to a budget model bills at the bottom of the band; a prompt routed to a frontier model bills at the top. We never stack a routing fee on top of the underlying model rate.

The worst case is published, too. No request can bill above $1.25 input / $3.95 output per 1M tokens. Budget at the ceiling — every request routed cheaper is savings you did not have to negotiate for.

USD per 1M tokens, as of 5 August 2026. The exact rate for each request appears on your invoice at the rate of the model that served it.

An honest fork

Pin a model, or let it choose

Pin a model when

  • Your workload is uniform — every prompt costs the same decision.
  • You are benchmarking or evaluating one specific model.
  • A contract or budget needs one guaranteed per-token rate.
Browse the catalog

Use Adaptive when

  • The workload is mixed — trivial prompts beside multi-step reasoning.
  • The right price-quality trade-off differs request by request.
  • You want failover between models without building it yourself.
Start building for free

Integration

Change two lines

OpenAI-compatible. Your existing client, your existing code.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.usbios.ai/v1",   # was api.openai.com
    api_key="sk-bios-...",
)

resp = client.chat.completions.create(
    model="bios-adaptive",           # was a pinned model id
    messages=[{"role": "user", "content": "Classify this ticket"}],
)

Prefer to choose models yourself? Browse the catalog