Muse Wonder: Opus 5.5 quality model that supports audio

Yesterday (US time), Meta released two new frontier models, Muse Wonder and its bigger sibling Muse Wonder Full (wonderful - get it?).

Meta pitches them as "personal and professional agents", built around two features:

• Active Memory, which carries context about you across conversations (see Memory).
• Active Focus, which sticks with a long task, works through blockers and checks in before important actions.

Although, I'm pretty sure these features aren't available through the API (yet?) (Meta).

Only Muse Wonder is available in the API. Muse Wonder Full needs a Muse Code subscription (the Power Usage or Maximum Usage plan), so it has no per-token price (Meta, 2026).

One cool thing is that these models support audio directly, alongside text, images, video and PDFs (Meta). I've always wondered when Claude models would support audio. I'm sure it's something they're looking at.

The model capability is very similar to Claude Opus 5.5, so I use that as the comparison for this article.

Specs Muse Wonder vs Opus 5.5

As of 13 October 2026 (Meta) (Meta) (Anthropic).

Pricing and Cache Settings

It's also priced exactly the same per token as Opus 5.5, down to the cache-read rate, so I compare the two throughout.

Prices in USD per million tokens, as of 13 October 2026.

Neither charges a long-context premium, so a nearly full 1M-token window costs the same per token as a short prompt (Meta) (Anthropic).

The caching works differently, though. Meta's caching is implicit. That is, it runs on every request, with no breakpoints to mark and no separate cache-write price listed. You can pass a prompt_cache_key to improve hit rates, or ask for 24-hour retention with prompt_cache_retention, but Meta describes the retention setting as "a hint, not a guarantee" (Meta).

On Opus you opt in with cache_control and pay 1.25x (5 minutes) or 2x (1 hour) the input price to write the cache (Anthropic).

Tokenisers

I counted the same six files I used for Claude Haiku 5.5 with each provider's token-counting endpoint (Anthropic's count_tokens and Meta's /v1/responses/input_tokens (Meta)) on 13 October 2026:

So Opus produces about 1.5 to 1.65 times as many tokens for the same text. At identical per-token prices, Muse Wonder is about 1.5 to 1.65 times cheaper for the same input, and output is likely similar. Muse Wonder's counts are within about 2% of GPT-6 Luna's from the Haiku article.

Cost Per Task

Artificial Analysis gives a better sense of what you actually pay. Running its Intelligence Index, both at max effort (ArtificialAnalysis) (ArtificialAnalysis):

So Opus scores higher, but uses twice the output tokens and costs about a third more per task. Artificial Analysis also measured Muse Wonder's time to first token at 19.27 seconds on Meta's API, so it's not fast to get going (ArtificialAnalysis).

Testing the audio

Classifying sounds

I took 25 five-second clips from ESC-50, a sound classification dataset with 50 labels like dog, rain, siren and church bells (Piczak, 2015). I sent each clip to the API with the full list of labels and asked it to pick one (Meta).

It got 20 of 25 right (80%). Guessing would get 2%, and human listeners in the original paper scored 81.3% (Piczak, 2015).

Clips from ESC-50 (CC BY-NC 3.0), originally from Freesound.

Three of the misses were "dog", the first label in the list. When I asked it to describe those clips, it often said "no recording was attached", even though the audio was sent. So sometimes it just doesn't hear the clip, and it doesn't tell you.

The 25 clips cost about US$0.69, mostly in thinking (raw results).

Transcribing two speakers

I also asked it to transcribe 75 seconds of a NotebookLM Audio Overview about Muse Wonder, starting a new line at each change of speaker.

It was basically perfect, with only one or two words off, and it even caught a quiet "Right, yeah" that WhisperX missed. Each run cost about US$0.05 (raw results).

Benchmarks

Meta's model card compares both Muse Wonder models against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, with every model at max reasoning effort (Meta, 2026). A selection of rows from its scorecard:

Even in Meta's own table, Opus 5.5 wins on computer use, terminal coding and ProgramBench. Muse Wonder wins on legal work, memory and long-running software tasks, though the biggest margins belong to Muse Wonder Full, which isn't on the API. Some of these benchmarks are Meta's own, and not every row used the same Coding Harness (Meta, 2026).

Overall, Opus 5.5 still looks like the stronger model at the same price per token, but Muse Wonder may be cheaper per task, and it's worth trying for long-running agent work.

Real example

As with Claude Opus 5.5 and GPT 6 Sol and Luna, I asked Muse Wonder at high effort to turn this article into a 3D gameshow with an N64 look, using the same prompt.

Meta's API is compatible with the OpenAI Python SDK, so the call is the same as the GPT 6 Sol example with a different base_url (Meta). The Notebook HTML helper saves the game and token usage, then renders it:

pip install openai ./notesbylex_notebook_html-0.1.2-py3-none-any.whl

Set MODEL_API_KEY in the environment used to launch Obsidian. Running the cell makes a paid request; the saved game plays without an API key or another model call.

import os
from openai import OpenAI
from notebook_html import read_article, render_openai_html

client = OpenAI(api_key=os.environ["MODEL_API_KEY"], base_url="https://api.meta.ai/v1")
article = read_article("muse-wonder.md", before="Real example")
with client.responses.stream(
    model="muse-wonder-1.0",
    max_output_tokens=128_000,
    reasoning={"effort": "high"},
    instructions="Return a complete, self-contained HTML document.",
    input=f"""
Turn this article into a 3D gameshow that looks like an N64 game.
Use around 8 questions that test understanding of the article, with a short
explanation after each answer.
Support phones and keyboards. Use inline CSS and JavaScript, no external assets,
network requests or browser storage, and no em dashes. Keep the code compact,
with one shared question renderer.

Article:
{article}
""",
) as stream:
    response = stream.get_final_response()
render_openai_html(response, "../_media/muse-wonder-quiz-high.html")

This game was generated on 13 October 2026 using muse-wonder-1.0 at high effort. It cost about US$0.12, calculated from the saved token usage and Meta's standard API prices (Meta):

There were no cache reads. Of the output tokens, 1,296 were reasoning tokens and 4,539 were the HTML response. Reasoning is billed at the output rate (Meta).

For comparison, the Claude Opus 5.5 used 49,136 tokens and cost about US$0.96, so this run was about 87% cheaper at the same per-token prices. Opus spent 36,150 tokens thinking; Muse Wonder spent 1,296. The GPT 6 Sol and Luna was cheaper again at about US$0.09. Each run used the same prompt at high effort, but with its own article as input.

Play the game in its own page.