How Much Do Tokens Cost for Meta’s Muse Wonder?

2026-11-01

Key Takeaways

Meta’s Muse Wonder costs $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens on the Meta Model API, where it is in preview as muse-wonder-1.0.
Reasoning tokens are billed at the output rate. The lowest effort setting is “minimal”; requests asking for no reasoning at all return an error.
Web search costs an extra $2.50 per 1,000 queries.
Caching is automatic with no fee for writing to the cache, and there is no surcharge for long prompts within the 1,048,576-token context window.
No batch discount is listed, and the cheap Contributor mode offered for Muse Spark does not apply to Muse Wonder.
Muse Wonder matches Claude Opus 5.5’s list price, while Meta’s own Muse Spark 1.3 costs $1.25 input and $4.25 output.
Muse Wonder Full is not sold by the token. It comes only with Muse Code Power ($50 a month) and Max ($200 a month) plans.

Muse Wonder – artistic impression. Image source: Meta

Tokens for Meta’s Muse Wonder cost $4.00 per million for input and $20.00 per million for output on the Meta Model API. Repeated input that hits the cache drops to $0.20 per million tokens, a 95% discount. Reasoning tokens count as output, and web searches carry a separate fee of $2.50 per 1,000 queries.

These are preview prices for the model Meta released on October 12, 2026.

The list price puts Muse Wonder level with Anthropic’s Claude Opus 5.5 and above OpenAI’s GPT-6.1 Sol. How much a project actually costs depends on how much the model reasons, how much of each prompt can be cached, and whether a subscription plan makes more sense than paying per token.

Muse Wonder API pricing table

Item	Price
Input tokens	$4.00 per 1 million
Cached input tokens	$0.20 per 1 million
Output tokens (including reasoning)	$20.00 per 1 million
Cache writes	No charge
Long-context surcharge	None
Web search	$2.50 per 1,000 queries
Batch discount	Not listed
Rate limits	3,000 requests and 4 million tokens per minute, per team

Muse Wonder accepts text, images, video, audio and PDF files and returns text. Its context window holds 1,048,576 tokens, and the same rate applies whether a prompt uses 5,000 tokens or close to a million.

How Muse Wonder billing works

Reasoning tokens count as output

Like most frontier models, Muse Wonder thinks before it answers, and every thinking token is billed at $20 per million. API responses report these separately in a reasoning_tokens field, which makes it easy to see how much of a bill comes from hidden reasoning. Effort settings run from minimal to max. Developers cannot switch reasoning off completely: a request set to “none” returns an HTTP 400 error, so minimal is the cheapest option.

Caching is automatic

Meta caches repeated prompt content automatically and does not charge to write it. That differs from Anthropic, which charges a premium to write to its cache before giving discounts on reads. For apps that resend the same system prompt, instructions or documents with every call, the 95% discount on cached input can cut input costs sharply.

Search adds a per-query fee

When Muse Wonder searches the web during a task, Meta adds $2.50 for every 1,000 queries. That works out to a quarter of a cent per search. Research agents that run dozens of searches per task should count this alongside token costs.

No cheap Contributor mode for Wonder

Meta offers a heavily discounted Contributor mode for Muse Spark 1.3, where it may train on the traffic in exchange for lower prices. That option is not available for Muse Wonder, so every Wonder request pays the full rate.

Muse Wonder vs other AI Models: token prices compared

Model	Input per 1M	Cached input per 1M	Output per 1M	Notes
Meta Muse Wonder	$4.00	$0.20	$20.00	Free cache writes; no batch discount listed
Claude Opus 5.5	$4.00	$0.20	$20.00	Cache writes $5 (5-minute) or $8 (1-hour); batch 50% off
Meta Muse Spark 1.3	$1.25	$0.15	$4.25	Contributor mode $0.10 input, $0.20 output
Gemini 4 Argon	$2.00 ($4.00 from Dec 1)	$0.10 ($0.20 from Dec 1)	$10.00 ($20.00 from Dec 1)	Batch and Flex 50% off
GPT-6.1 Sol	$2.00	$0.10	$10.00	Batch and Flex 50% off
GPT-6 Astra	$10.00	—	$50.00	Premium frontier tier

Claude Opus 5.5 shares Muse Wonder’s headline rates, but the fine print differs. Anthropic’s batch tier halves the price to $2 input and $10 output for work that can wait, and its fast mode costs $8 input and $40 output. US-only data processing adds a 1.1 times multiplier.

Meta’s free cache writes favor short, repeated tasks, while Anthropic’s batch discount favors large overnight jobs. For a broader look at pricing gaps between labs, our analysis of whether Grok 4.5 is five times cheaper than Claude Opus shows how much these details matter.

What Muse Wonder costs in practice: three examples

A busy day of chat and assistance

Suppose an internal assistant processes 2 million input tokens in a day, with 60% of that served from cache, and produces 500,000 output tokens. The 800,000 uncached input tokens cost $3.20, the 1.2 million cached tokens cost $0.24, and the output costs $10.00. The day’s bill comes to about $13.44, and output accounts for three-quarters of it.

One deep research task

A research run that reads 300,000 tokens of material, 250,000 of them cached, writes 40,000 tokens including reasoning, and performs 20 web searches costs about $1.10: $0.20 for fresh input, $0.05 for cached input, $0.80 for output and $0.05 for search.

An overnight processing job

A large job with 50 million input tokens and 4 million output tokens, with no caching, costs $280 on Muse Wonder. The same job on Claude Opus 5.5’s batch tier would cost $140, and on Muse Spark 1.3 about $79.50. For bulk work where top-tier reasoning is not essential, Wonder is the expensive choice.

Muse Code subscription plans: an alternative to paying per token

Developers who mainly use Muse Wonder for coding can skip the API and subscribe to Muse Code, Meta’s coding agent.

Plan	Monthly price	Usage allowance	Muse Wonder Full
Everyday	$5	About 10 to 50 prompts per 5 hours	No
High	$15	5 times Everyday	No
Power	$50	20 times Everyday	Yes
Max	$200	80 times Everyday	Yes

Two limits apply. A subscription key only works inside the Muse Code command-line tool, so it cannot power a separate app. And Wonder Full, the larger and more capable configuration of the model, is only available on the Power and Max plans; selecting maximum effort on the standard model does not unlock it. Plans require users to be 18 or older, are offered in selected regions, and are non-refundable.

How to keep Muse Wonder costs under control

Output tokens drive most of the bill, so the effort setting is the first lever. Running routine steps at minimal or low effort and saving high effort for hard problems can cut costs substantially.

Placing fixed instructions and reference documents at the very start of each prompt lets the automatic cache do its work. Capping output length and limiting the number of search calls per task keep the other variable charges predictable.

Mixing models helps too. Muse Spark 1.3 handles many everyday tasks at around a fifth of Wonder’s output price, and teams can reserve Wonder for the agent steps that need its memory and focus features.

Lower-priced rivals, including several cheap Chinese AI models that rival the big players, are also worth testing for bulk work. Understanding what a token is and how it is counted makes it easier to estimate any of these bills in advance, and comparing Wonder with GPT-6 Astra’s token usage and cost shows how widely frontier pricing varies. Meta’s wider model plans are covered in our preview of what to expect from Meta AI in 2026.

If you are interested in this topic, we suggest you check our articles:

Sources: eesel AI – Muse Wonder pricing, eesel AI – Muse Wonder, Meta – Muse Wonder, Anthropic – Claude Opus 5.5, eesel AI – Gemini 4 Argon pricing

Written by Alius Noreika