Muse Wonder on Meta's Model API: what $4/$20 buys next to Muse Spark and Gemini 4 Argon

Meta released two new models on 12 October: Muse Wonder and Muse Wonder Full, which it calls its most capable models so far. Only one of them is available to developers. Muse Wonder is on the Meta Model API as muse-wonder-1.0. Wonder Full is limited to Muse Code subscribers. This post covers what developers can actually call, what it costs, and how it compares with Meta’s cheaper Muse Spark and with Google’s Gemini 4 Argon, which reached the Gemini API a day later.

What Meta shipped

According to Meta’s research post and newsroom announcement:

• Muse Wonder is available through the Meta Model API and Muse Code.
• Muse Wonder Full is available only to Muse Code subscribers. Meta lists no API price for it.
• Both target personal and professional agent tasks, with law, finance, health and coding as the named domains.
• Meta highlights two behaviours: Active Memory, which carries context across conversations, and Active Focus, which keeps working towards a goal, tries other approaches when blocked and checks in before important steps.
• Wonder will also “soon power experiences” in the Muse consumer app, rolling out to subscribers first.

The Model API docs say plainly that Muse Wonder is in preview, so expect changes before you build anything critical on it.

The spec sheet

From Meta’s models page and the Muse Wonder model page:

• Context window: 1,048,576 tokens.
• Input: text, images, video, audio and PDF. Output: text only.
• Reasoning effort: adjustable from minimal to max.
• Features: tool calling with parallel calls, structured output, search grounding, computer use (operating browsers, desktop and mobile apps) and delegating work to sub-agents. The cheaper Muse Spark 1.3 has the first three but not computer use or sub-agent delegation.
• Max output tokens: Meta’s pages don’t state a limit.

Pricing and limits

From Meta’s pricing and rate limits page, per million tokens:

	Muse Wonder 1.0	Muse Spark 1.3 (Standard)	Muse Spark 1.3 (Contributor)
Input	$4.00	$1.25	$0.10
Cached input	$0.20	$0.15	$0.002
Output	$20.00	$4.25	$0.20
Web search tool	$2.50 / 1,000 queries	$2.50 / 1,000 queries	not stated
Requests per minute	3,000	3,000	150
Tokens per minute	4M	4M	3M

Three points stand out:

1. No long-context surcharge. Meta says the same rate applies however much of the context window you fill. Rate limits apply per team, not per API key.
2. No cheap Contributor mode for Wonder. On Muse Spark, Contributor mode cuts the price sharply, but it requires permission for Meta to use your prompts, files, code and responses to train future models. The pricing page shows Wonder in Standard mode only. For most business data that’s the safer default anyway: don’t send client or personal data through a mode that allows training on it.
3. Wonder costs about 3–5× Spark. Input is 3.2× and output about 4.7× the Spark price. Before you move to Wonder, test whether your task needs it. Many extraction, classification and summarisation jobs won’t.

Switching is mostly a config change

The overview docs say the Model API is drop-in compatible with the OpenAI SDK, the Anthropic SDK and OpenAI-compatible agent CLIs. It supports the Responses, Chat Completions and Messages APIs. You point your client at https://api.meta.ai/v1, supply a Meta key and keep the rest of your code. That makes a side-by-side trial cheap: run the same eval set through your current provider and Wonder by changing only the base URL and model name.

Muse Wonder vs Gemini 4 Argon

Google opened Gemini 4 Argon to paid API customers on 13 October, so many teams will evaluate the two together. Here is how they compare on paper, using each vendor’s own pages:

	Muse Wonder 1.0 (Meta)	Gemini 4 Argon (Google)
Status	Preview (Meta docs)	Not stated on Google’s launch pages
Input / output per 1M tokens	$4 / $20	$2 / $10 until 30 Nov; $4 / $20 from 1 Dec (pricing)
Cached input per 1M tokens	$0.20	$0.10 until 30 Nov; $0.20 from 1 Dec (implicit caching only)
Discounted async tiers	None listed	Batch and Flex at half the Standard price
Free tier	None listed	Not on the Free tier
Context / long output	1,048,576-token context	Output up to 1,048,576 tokens minus input, through background long decode
SDK compatibility	OpenAI and Anthropic SDKs	Google’s Gemini API

From December the list prices match. Until then Argon costs half as much, and its Batch and Flex tiers halve the price again for work that can wait. Wonder’s advantages are the drop-in SDK compatibility, no long-context surcharge and computer use built into the model.

On quality, Meta’s own pages claim state-of-the-art results on agent tasks but don’t publish a score table in the text we could read. VentureBeat reports Artificial Analysis Intelligence Index scores of 52 for Wonder, 56 for Wonder Full and 53 for Gemini 4 Argon. That puts the API version of Wonder roughly level with Argon, behind the subscription-only Wonder Full. Treat those numbers as indicative until you’ve run your own evals.

Safety and data questions to ask first

Meta says Wonder was developed under its Superintelligence Scaling Framework and that a fuller safety report is “available soon” (research post). Until it’s out, there is little public safety documentation for Wonder. If you work in a regulated area (health, lending, legal), keep a human review step on anything Wonder produces for customers. Also confirm with Meta which region processes your data, because none of the Model API pages we read say.

Bottom line

• Try Wonder if you already use the OpenAI or Anthropic SDKs and want a second frontier option for agentic or computer-use work. The trial is a two-line change.
• Stay on Spark for high-volume, simpler tasks. It costs a fraction of Wonder’s price on the same API.
• For a cost-driven pick before December, Gemini 4 Argon’s introductory and batch prices are lower. After 1 December the decision comes down to your own eval results, not list price.