BestCodes for AgentOne

Posted on Oct 12 Originally published at agent-one.dev

Meta Muse Wonder: Benchmarks, Pricing, API, and How to Use It

#musewonder #meta #aimodels #guide

Heads up: This article was written by AgentOne Research.

Meta released Muse Wonder today. It's the first Meta model in a while that's actually competitive with the top models from Anthropic and OpenAI.

I went through Meta's announcement, the model card, the API docs, and the early Artificial Analysis results. Below is what it is, how it scores, what it costs, and how to start using it. If you just want the API details, jump to setup.

What is Muse Wonder?

Muse Wonder is Meta's new frontier model, released on October 12, 2026. It replaces Muse Spark, which came out in September. Meta built it for agent work, meaning long tasks where the model uses tools, works across several apps, and keeps going until the job is done. Meta also put a lot of focus on law, finance, health, and coding.¹

It comes in two versions. Muse Wonder (muse-wonder-1.0) is the one you can actually use. It's on the Meta Model API, OpenRouter, and Muse Code.

Muse Wonder Full is the bigger model, and most of the impressive benchmark numbers are from it. Right now you can only use it inside Muse Code on the Power Usage and Maximum Usage plans. There's no API for it yet.²

It's also not open source. Meta used to release Llama weights for anyone to download, but Muse Wonder is closed, like Claude or GPT.

Specs

Spec	Muse Wonder
API model ID	`muse-wonder-1.0` (OpenRouter: `meta/muse-wonder`)
Released	October 12, 2026
Context window	1,048,576 tokens
Input	Text, images, video, audio, PDF
Output	Text
Reasoning effort	`minimal`, `low`, `medium`, `high`, `xhigh`, `max`
Tools	Tool calling, structured output, search grounding, computer use
API formats	OpenAI Responses, OpenAI Chat Completions, Anthropic Messages
Base URL	`https://api.meta.ai/v1`
Price (in / out)	$4.00 / $20.00 per 1M tokens
Rate limits	3,000 requests and 4M tokens per minute, per team
Open weights	No

Active Memory and Active Focus

Meta is marketing two features with Muse Wonder. Both are about how the model was trained, so there's nothing to turn on.

Active Memory is about holding onto context. In a long conversation, the model should remember the details that matter and bring them back when they become relevant. If the app it's running in has a memory system, that applies across conversations too. Meta tests this with an internal benchmark called MemoryBench, which hides facts in 64K to 100K token conversations, adds distractions, and changes some facts halfway through. Both Muse Wonder versions score about 85.5, which is higher than every other model Meta tested.³

Active Focus is about not giving up. When something fails, the model is supposed to try another way, work around the problem, and ask you when it's actually unsure. Meta also says it checks in before doing anything important.¹

I like the idea, but it has a downside in an agent. A model that won't give up will also keep trying things you might not want it to try. Meta's own model card recommends asking for human confirmation before any action that can't be undone, and I'd follow that.³

Artificial Analysis results

Artificial Analysis is the independent benchmark most people check first, so I'll start there. Their results went up the same day as the launch.⁴

Muse Wonder Full scores 56, tied with Claude Sonnet 5.5 and behind only Claude Opus 5.5 at 58. It beats Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon, which all score 53.

The regular Muse Wonder scores 52, which puts it #9 out of the 219 models Artificial Analysis tracks. That's one point behind Fable 5.1 and GPT-6 Astra, and tied with GPT-6.1 Sol. It's also a big jump from Muse Spark 1.3, which scored 48.

Artificial Analysis also found some specific strengths and weaknesses for Wonder Full:⁵

It's #1 on CritPt (physics reasoning) and GDP.pdf (professional document work)
It's #2 on GDPval-AA, and #3 on AA-Briefcase, AutomationBench-AA, and SciCode
Its general knowledge is weak. On AA-Omniscience it scores 30, compared with 46 for Opus 5.5 and 43 for GPT-6 Astra. It does have a very low hallucination rate, so when it doesn't know something, it usually says so instead of making things up.

The cost-per-task column is the important part, and I'll come back to it in the pricing section.

Meta's benchmarks

Meta published its own comparison table in the model card. Everything is at max reasoning effort, against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5.³

Personal and professional agents

This is where Muse Wonder does best. Wonder Full leads Meta's Personal Superintelligence Index (59.1), Vals Finance Agent V2 (59.1), and AutomationBench-AA (45.4). The biggest lead is on Harvey LAB-AA, a legal benchmark, where Wonder Full scores 13.1 and Opus 5.5 scores 4.2. Those are low numbers because the test only counts a task as passed if every requirement is met and there are no hallucinations.

Regular Muse Wonder beats Fable 5.1 on most of these and has the top MemoryBench score (85.6).

Opus 5.5 still wins on HealthBench Pro and JobBench, and GPT-6 Astra wins on GDP.pdf. On OSWorld V2.1, which tests using a computer through screenshots, Opus 5.5 leads too, with 81.8 compared with 78.4 for Wonder Full and 72.6 for Wonder.

Coding

Coding results are mixed. Muse Wonder is very good at long projects. It leads DeepSWE 1.1 (79.0 for Full, 78.8 for regular) and SWE-Marathon (57.5 for Full, 53.1 for regular), which tests very long engineering tasks.

On shorter, self-contained coding tasks, Opus 5.5 is ahead. It's well ahead on ProgramBench (70.5 vs 59.0) and slightly ahead on Terminal-Bench 4.0 (59.6 vs 57.1).

What to keep in mind

Meta ran some of these tests itself. Some numbers come from third parties like Artificial Analysis, Vals AI, and Datacurve, but the Personal Superintelligence Index, MemoryBench, and FocusBench are Meta's own internal benchmarks. I'd wait for someone else to test those.

The SWE-Marathon row also isn't a fair comparison. Meta notes that GPT-6 Astra and Fable 5.1 were tested through Codex and Claude Code, while Wonder Full ran at xhigh effort, so the models weren't tested the same way.

And most of the wins belong to Wonder Full. If you're using the API, you're getting the regular Muse Wonder, which is about 4 points lower on the Artificial Analysis index.

Muse Wonder pricing

Model	Input (per 1M)	Output (per 1M)	Cached input (per 1M)	Cost per AA task
Muse Wonder	$4.00	$20.00	$0.20	$4.52
Claude Fable 5.1	$10.00	$50.00	$0.25	$7.63
GPT-6 Astra	$10.00	$50.00	$1.00	$3.26
Muse Spark 1.3	$1.25	$4.25	$0.15	$1.60

Token prices for Muse models are from Meta's pricing page.⁶ Cost per task is what Artificial Analysis measured running its index at each model's max setting.⁴ For comparison, Claude Opus 5.5 costs $5.98 per task and GPT-6.1 Sol costs $0.72.

Per token, Muse Wonder is 60% cheaper than Fable 5.1 and GPT-6 Astra.

The catch is that Muse Wonder uses a lot of tokens. Artificial Analysis says it generated 130 million output tokens running their index, compared with a median of 85 million. Once you account for that, it costs $4.52 per task. That's still cheaper than Fable 5.1 and Opus 5.5, but it's more expensive than GPT-6 Astra, and about six times the price of GPT-6.1 Sol, which scores the same 52.

So if you're comparing it with Claude, Muse Wonder is cheaper. If you're comparing it with OpenAI's models, it isn't.

A few other pricing details:

Contributor mode (cheaper tokens if you let Meta train on your data) is available for Muse Spark but not for Muse Wonder.⁶
Meta's built-in web_search tool costs $2.50 per 1,000 searches on top of token costs.
In the Muse app, Muse Wonder will be included in the $20 and $100 monthly plans.⁷

How to use Muse Wonder

With the API

The Meta Model API is OpenAI-compatible, so most existing code works if you change the base URL and model name. Create a key under API keys at dev.meta.ai. Muse Wonder doesn't have a contributor tier, so you'll need billing set up. Availability also depends on your country.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key="YOUR_META_API_KEY",
)

response = client.chat.completions.create(
    model="muse-wonder-1.0",
    reasoning_effort="high",
    messages=[{"role": "user", "content": "Summarize this lease and flag anything unusual."}],
)

print(response.choices[0].message.content)

Some things that are useful to know:

reasoning_effort accepts minimal through max. Muse Wonder doesn't support none, and sending it returns a 400 error. If you leave it out, the model picks a level on its own.⁸
On Chat Completions, the model's reasoning doesn't carry over from one turn to the next. If you want it to, use the Responses API with previous_response_id.
The API also accepts the Anthropic Messages format, so tools built for Claude can use it too.

With OpenRouter

Muse Wonder is on OpenRouter as meta/muse-wonder, so it works anywhere you already use an OpenRouter key.⁹

In AgentOne

If you'd rather use it in a desktop app with tools, AgentOne can run it. Full disclosure, I make AgentOne. It doesn't have a built-in Meta provider yet, so you add it as a custom one.

Open Settings > Providers. Under Custom, click Add Provider and choose OpenAI Compatible.
Set Name to Meta, Base URL to https://api.meta.ai/v1, and paste your API key.
Under Models, click Auto to load Meta's models, or click Add Model and enter muse-wonder-1.0 with a context window of 1048576. Turn on Supports Tools and Supports Images.
Click Add Provider, then pick Muse Wonder from the model selector in a chat.

If you already use OpenRouter in AgentOne, you can open the OpenRouter provider settings and add meta/muse-wonder with Add Model Override.

AgentOne doesn't have a reasoning effort setting yet, so Muse Wonder will choose its own level.

Limitations

Meta lists these in the model card:³

It's weak at visual and spatial reasoning, especially anything 3D.
It still makes mistakes on multi-step problems, especially ones that don't look like its training data.
Results depend a lot on the app, prompts, and tools you use it with, so you might not see the same numbers as Meta.
Not every language was tested.
It isn't meant for anyone under 18.

Artificial Analysis also found that its general knowledge is weaker than the other top models, as I mentioned above. Meta's full safety report hasn't been published yet.

Should you use it?

If your work involves long agent tasks, legal or financial documents, or automating work across several apps, Muse Wonder is worth trying. It does best on exactly those benchmarks, and per token it costs a lot less than Claude.

If you mostly want the smartest model available, Claude Opus 5.5 is still ahead. If you mostly care about cost, GPT-6.1 Sol gets the same Artificial Analysis score for much less money per task.

The model I'd be most interested in is Wonder Full. It scores right next to Opus 5.5, but until Meta puts it on the API, most people won't be able to use it.

Footnotes

Meta AI Research, "Muse Wonder," October 12, 2026. https://research.meta.ai/blog/muse-wonder ↩
Meta Model API, "Models." https://dev.meta.ai/docs/models ↩
Meta, "Muse Wonder Model Card," October 12, 2026. https://research.meta.ai/static/muse-wonder-model-card ↩
Artificial Analysis, LLM Leaderboard and "Muse Wonder (Max)" model page, accessed October 12, 2026. https://artificialanalysis.ai/models/muse-wonder ↩
Artificial Analysis on X, via Techmeme, October 12, 2026. https://www.techmeme.com/261012/p36 ↩
Meta Model API, "Pricing and rate limits." https://dev.meta.ai/docs/pricing-rate-limits ↩
CNBC, "Meta debuts Muse Wonder AI model in latest test of users' willingness to pay," October 12, 2026. https://www.cnbc.com/2026/10/12/meta-muse-wonder-ai-model-agent.html ↩
Meta Model API, "Reasoning." https://dev.meta.ai/docs/reasoning ↩
OpenRouter, "Muse Wonder." https://openrouter.ai/meta/muse-wonder ↩

DEV Community