Get an API key: Meta Model API

new

The Meta Model API serves Meta's Muse Spark reasoning models and Muse Image on one pay-as-you-go key. Every Muse Spark reply reasons first, so the first word arrives a few seconds after you send.

Meta Model API account

Accounts live at dev.meta.ai. Billing is pay-as-you-go per token, charged at a rising payment threshold and monthly; the pricing page is the source of record. Keys, usage and billing belong to a team, and every key in the team shares the team's rate limit (the per-minute ceiling on requests and tokens).

Creating a key

1. Open dev.meta.ai/api-keys - the empty key field links there as get a Key.
2. Create a key and copy it: it is shown once.

Keys start with LLM_, followed by digits and a secret. The field accepts that prefix at 20 characters or more, so a wrong key of the right shape passes the field and is refused upstream with Unauthorized.

Add it to Big-AGI

Open Models (Ctrl + Shift + M) and press More Services if the Setup AI Models wizard appears. Then Add -> Meta Model API, paste into Meta Model API Key, and press Models - the list filling in below is your confirmation (add and manage your keys).

Advanced carries an API Host field for a proxy or a gateway of your own: give it a bare origin such as https://api.meta.ai, since Big-AGI appends the API paths itself.

Standard and Contributor tiers

Each Muse Spark version is served twice. Standard does not use your prompts and completions for training. Contributor (the -contributor ids) is heavily discounted in exchange for Meta training future models on your prompts and completions, with lower rate limits. Same model, same capabilities. Big-AGI ships the Contributor models hidden; unhide them from the model list to opt in.

What we measured, September 2026

Snapshot of 2026-09-02 from a US connection: single runs of short prompts through Big-AGI's own pipeline, not benchmarks, and prices read from the pricing page the same day.

Model	Context	$/1M in	$/1M out	Input
Muse Spark 1.3, 1.2, 1.1	1M	$1.25 (cache hits $0.15)	$4.25	text, images
Muse Spark 1.3 and 1.2 Contributor	1M	$0.10 (cache hits $0.002)	$0.20	text, images
Muse Image 1.0	-	$0.01 per image	-	text, reference images; one image out

Muse Spark reasons before every answer: a two-word reply spent about 340 reasoning tokens and 9 seconds at the default effort, first token after 2.2 s. The Reasoning Effort control runs from minimal to xhigh, and there is no off. Web search is a per-model switch and adds $2.50 per 1,000 searches to the token bill. Muse Image returns one WebP image per request and runs without streaming. Standard rate limits are 3,000 requests and 4M tokens per minute per team; Contributor is documented at 100 requests per minute. Video, PDF and audio input are served by the API but not yet carried by Big-AGI.

Capping the spend

Usage and billing live at dev.meta.ai/usage and dev.meta.ai/billing. No spend-cap control is recorded here, so the enforceable ceiling is whatever the billing page offers. The per-message cost Big-AGI shows is an estimate from the published rates, never a limit.

Direct Connection

Available: Meta's API accepts browser-direct requests, so the toggle appears under Advanced once the key is in the field. It starts off. On that path the browser cannot read Meta's rate-limit headers.

Using a server-side key

Self-hosters set METAAI_API_KEY in the server environment, and METAAI_API_HOST to change the endpoint. No green check appears and the field stays empty, but the server supplies the key on every request: add the service by hand, leave Meta Model API Key empty and press Models (keys held by a server you operate).

When this doesn't work

What you see	What it means	What to do
Unauthorized after pressing Models	The key is wrong, revoked, or missing on the server: Meta returns the same error for all three.	Recreate the key at dev.meta.ai/api-keys: key and access errors
A long pause, then the answer	Muse Spark reasons first, and the default effort is high.	Lower Reasoning Effort in the model parameters
The reply ends early or comes back empty	The response-token cap was spent on reasoning.	Raise the output tokens, or lower the effort
Rate limit exceeded	The team's per-minute limit is shared by every key, and the Contributor limits are far lower.	Wait a minute, or use the Standard model: key and access errors

Related

• Connect an AI service
• Direct Connection
• What your provider does with your text
• Keys held by a server you operate