API Reference Overview

The EmpirioLabs API speaks OpenAI- and Anthropic-compatible request shapes so any client that talks to those providers works against https://api.empiriolabs.ai unchanged. All requests require a bearer token (Authorization: Bearer $EMPIRIOLABS_API_KEY or the Anthropic-style x-api-key header).

Endpoints

Models as the source of truth

GET /v1/models is the canonical schema. Every model returns:

• pricing_rows[] — per-token, per-image, per-second, or per-call rates with any cache or volume discounts applied.
• supported_parameters[] — name + type + default + range for every parameter the model accepts (including conditional parameters that only apply when another flag is set).
• capabilities — flags for streaming, system_prompt, tool_calling, vision, audio_input, video_input, web_search, thinking, citations, etc.
• input_modalities / output_modalities — distinct lists so multimodal models that accept video as input but only emit text are not mistaken for video generators.

$
curl "https://api.empiriolabs.ai/v1/models" \
>
  -H "Authorization: Bearer $EMPIRIOLABS_API_KEY"

Universal disable_formatting flag

Every chat, search, and media endpoint accepts a disable_formatting=true flag (also accepted as raw=true, passthrough=true, or raw_response=true). When set, the worker skips EmpirioLabs server-side formatting (Markdown “Thinking…” headers, citation list reformatting, code-block wrapping, etc.) and returns the upstream payload shape verbatim. Use this when:

• You render responses yourself and want full control of the markdown.
• You feed the response into a downstream pipeline that expects the raw upstream shape.
• You are debugging why a response looks different in the playground vs your app.

Tool calling

Tool / function calling propagates straight through. Pass standard OpenAI tools[] and tool_choice on /v1/chat/completions; the response carries message.tool_calls[] and a downstream role: "tool" message round-trips correctly. Anthropic-style tool blocks work the same way on /v1/messages.

Long-running requests

The platform’s inbound HTTP timeout is 15 minutes (the maximum a Railway edge connection holds open). Anything that needs longer must use the asynchronous POST /v1/jobs/... family:

• /v1/audio/transcriptions returns a job_id immediately.
• /v1/images/generations and /v1/videos/generations always create a job (image jobs typically complete in seconds; video jobs in 1-15 minutes).
• Poll GET /v1/jobs/{job_id} every 2-10 seconds. Job state is retained for up to 1 hour after completion.

If a synchronous chat completion needs more than ~14 minutes of compute (deep reasoning, multi-step research), set stream=true so partial output is returned as it arrives.

Generated media retention

All generated images, videos, audio, and stored transcript files live on the EmpirioLabs media CDN at https://media.empiriolabs.ai. URLs returned by the API are signed with a 7-day expiry. The underlying objects are deleted after 30 days. Save anything you need to keep — both at the URL level and the file level — within that window.

Errors

Errors come back in OpenAI’s shape ({"error": {"message", "type", "code"}}) on chat/responses endpoints, and Anthropic’s shape on /v1/messages. Errors never leak the upstream provider name, internal hostnames, or provider request IDs — those are stripped by the gateway and worker layers before the response reaches you.

How to use the generated reference

• Open API Reference in the top navigation tab.
• Inspect endpoint details, required fields, and example payloads.
• Use the example responses to align your integration with the documented contract.
• Use AI Agent Access when you want an AI coding assistant to read the docs through llms.txt or machine-readable API schemas.