Model calls

A deployed app calls OpenAI models through floo with the stock OpenAI SDK. There is no key to create or store: the app proves who it is with a short-lived token for its own identity, and floo bills the tokens to your org’s credits. Model calls are in early access. Ask us to turn them on for your org.

Call a model

Every deploy sets FLOO_AI_BASE_URL. Fetch an identity token for that URL and pass it as the API key.

Python

import os
import requests
from openai import OpenAI

base_url = os.environ["FLOO_AI_BASE_URL"]
token = requests.get(
    "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity",
    params={"audience": base_url},
    headers={"Metadata-Flavor": "Google"},
).text

client = OpenAI(base_url=base_url, api_key=token)
response = client.responses.create(model="openai/gpt-6-luna", input="Summarize this ticket: ...")
print(response.output_text)

Node

import OpenAI from "openai";

const baseURL = process.env.FLOO_AI_BASE_URL;
const tokenUrl =
  "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity" +
  `?audience=${encodeURIComponent(baseURL)}`;
const token = await (await fetch(tokenUrl, { headers: { "Metadata-Flavor": "Google" } })).text();

const client = new OpenAI({ baseURL, apiKey: token });
const response = await client.responses.create({ model: "openai/gpt-6-luna", input: "Summarize this ticket: ..." });
console.log(response.output_text);

The token lasts one hour. Fetch a new one for each job, or cache it for up to 50 minutes. The token works only inside a deployed app. Running the app on your laptop is not supported yet.

Models

Model	Use it for	Endpoints
`openai/gpt-6-astra`	The hardest reasoning and coding work	`responses`
`openai/gpt-6.1-sol`	Complex work at a fifth of Astra’s price	`responses`
`openai/gpt-6.1-terra`	Vision and writing	`responses`
`openai/gpt-6-luna`	Cheap, high-volume tasks	`responses`, `decisions`

GET $FLOO_AI_BASE_URL/models lists the same ids. Responses report the floo id in model. decisions is OpenAI’s Decisions API: typed answers (yes/no probability, a choice, or a score) about some input, faster than responses. Its output tokens are free.

Pricing

Tokens bill at OpenAI’s list price plus a 10% processing fee, from your credits, as each call returns. Rates per million tokens are on the pricing page under Models. A prompt over 272K input tokens bills at the long-context rates. Calls from your prod environment bill to prod. Calls from dev and preview environments bill to dev.

What floo does not pass through

One OpenAI account serves every floo app, so nothing a request sends may be stored there. These requests are refused with 400:

Request	Do this instead
`previous_response_id`, `conversation`, `background`, a stored `prompt`	Send the full input on each request
Hosted tools: web search, file search, code interpreter, image generation, computer use, MCP	Use `function` or `custom` tools and run them in your app
`file_id` anywhere in the input	Send images and files inline or by URL
`stream: true`	Wait for the full response; streaming is not available yet
A `service_tier` other than `default` or `auto`	Omit it
`store` on `decisions`	Omit it

floo sets store: false on every responses call and scopes safety_identifier and prompt_cache_key to your app.

Limits

600 model calls per minute per org.
8 model calls running at once per org.
A call starts only while your org has credits. A call that starts with credits left is billed in full when it returns.

Errors

Status	Code	Meaning
401	`UNAUTHORIZED`	The token is missing, expired, or not for a floo app. Fetch a new one with `FLOO_AI_BASE_URL` as the audience.
402	`SPEND_AUTHORITY_EXHAUSTED`	Your org is out of credits.
403	`AI_NOT_ENABLED`	Model calls are not on for your org yet.
429	`RATE_LIMIT_EXCEEDED`, `AI_ORG_CONCURRENCY_LIMIT`, `AI_ENDPOINT_BUSY`	Too many calls. Retry after `Retry-After` or shortly.
502	`MODEL_PROVIDER_UNAVAILABLE`	The model provider failed or did not respond. Nothing was billed.
503	`AI_NOT_CONFIGURED`, `CALLER_IDENTITY_UNAVAILABLE`	floo cannot serve model calls right now. Retry.

Other errors from the model come back with OpenAI’s status and an error object, and are not billed.

Get started

Build with your stack

Deploy from your agent

Configure your app

Ship and release

Data

Domains and networking

App users and access

Operate

Team and billing

For agents

Model calls

Call a model

Models

Pricing

What floo does not pass through

Limits

Errors

​Call a model

​Models

​Pricing

​What floo does not pass through

​Limits

​Errors

Call a model

Models

Pricing

What floo does not pass through

Limits

Errors