OpenAI compatibility

Bayle’s API is shaped like OpenAI’s, so an OpenAI SDK works against it once you change the base URL and the key. Error responses use OpenAI’s envelope and its type vocabulary too, which means an SDK’s exception mapping keeps working without changes.

Changing the base URL

Python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["BAYLE_API_KEY"],
    base_url="https://api.bayle.ai",
)

JavaScript

import OpenAI from 'openai'

const client = new OpenAI({
  apiKey: process.env.BAYLE_API_KEY,
  baseURL: 'https://api.bayle.ai',
})

Go

import (
    "os"

    "github.com/openai/openai-go"
    "github.com/openai/openai-go/option"
)

client := openai.NewClient(
    option.WithAPIKey(os.Getenv("BAYLE_API_KEY")),
    option.WithBaseURL("https://api.bayle.ai"),
)

Environment only

# Many tools read these directly, so no code change is needed at all.
export OPENAI_API_KEY="$BAYLE_API_KEY"
export OPENAI_BASE_URL="https://api.bayle.ai"

The last tab is worth knowing about: agentic CLIs and editor integrations that speak to an OpenAI-compatible endpoint usually pick up OPENAI_BASE_URL and OPENAI_API_KEY without any configuration of their own.

Where behaviour differs

Three of these will surprise a client written against OpenAI’s defaults. All three are on the images endpoint.

	OpenAI	Bayle
Images per request	n up to 10, model depending	At most 1; a higher n is rejected
response_format: "url"	A fetchable https:// link	A base64 data: URI in the url field
response_format omitted	Defaults to url	Returns b64_json, with no url field at all

Caution

The third row is the one that costs people time. Reading data[0].url after omitting response_format, which works against OpenAI, finds nothing here. Either set response_format: "url" explicitly, or read data[0].b64_json. See Images.

Extra fields in responses

Bayle returns a few fields beyond OpenAI’s published schema. An SDK ignores what it doesn’t recognize, so these are additive rather than breaking:

• cost, an object breaking out what the request cost by the balance it was drawn from.
• service_tier, the tier that served the request.
• reasoning_content and thought_signature inside a streamed delta, for reasoning-capable models. Both are null for models that don’t reason.

What Bayle does not serve

The API surface is exactly the eight operations in the reference. There is no assistants API, no files API, no fine-tuning, no moderations endpoint, and no audio endpoints. A call to anything not listed there will not find a route.

Beyond the OpenAI surface

Video generation has no equivalent in OpenAI’s API, so it doesn’t come along with an SDK. It’s a queue you poll, over four endpoints, and it is covered in the video guide.