Meta Releases Muse Wonder Models at $4/$20 per Million Tokens

Meta has released Muse Wonder and Muse Wonder Full, its most capable models so far, with Muse Wonder priced at $4 per million input tokens and $20 per million output tokens and a 1M-token context window. Muse Wonder is live now on the Meta Model API and Muse Code; Muse Wonder Full is restricted to paid Muse Code plans starting at $50 per month.

The release targets long-running agent work: coding, legal and finance tasks, and computer use. Meta positions Muse Wonder as a frontier model built for personal and professional AI agents. It follows Muse Spark on Meta's scaling ladder, and Meta states the company trained it at its Prometheus data center, with gains across pretraining and reinforcement learning. Meta credits a larger base model for better results on unfamiliar tasks.

The stakes are straightforward. Anthropic and OpenAI currently anchor the paid agent market with Claude Opus 5.5 and GPT-6 Astra, both priced at $10/$50 per million tokens. Meta enters the same tier at half the input and output price, with four input modalities — text, image, video and audio — against text and image for both competitors.

What defines the Muse Wonder release?

Two named behaviors carry the launch:

• Active Memory: carries user context across conversations and pulls back the details a task needs.
• Active Focus: keeps working on hard tasks, tries alternatives, and checks in before important actions.

For developers, Muse Wonder plans long tasks and delegates to sub-agents. It also operates browsers, desktop and mobile apps. Both Muse Wonder variants run on Meta's cloud only, via the API or the Muse Code CLI on macOS, Linux and Windows.

Neither model ships with disclosed parameters. Neither offers open weights. Meta has not published a full safety report for the release.

How does Muse Wonder perform on benchmarks?

All scores come from Meta's published table at max reasoning effort. Several rows use third-party sources such as Artificial Analysis, Vals AI and Datacurve. Muse Wonder Full leads 6 of 14 rows in the table. It trails Claude Opus 5.5 on several others.

Benchmark	Muse Wonder Full	Muse Wonder	Claude Fable 5.1	GPT-6 Astra	Claude Opus 5.5
Personal Superintelligence Index (internal)	59.1	54.3	49.9	54.5	55.2
Harvey LAB-AA	13.1	10.0	6.4	8.6	4.2
Vals Finance Agent V2	59.1	56.6	58.9	53.5	58.6
AutomationBench-AA	45.4	39.1	32.1	41.6	41.2
OSWorld V2.1 (partial)	78.4	72.6	80.7	75.8	81.8
MemoryBench (internal)	85.5	85.6	79.0	81.4	81.5
Terminal-Bench 4.0-AA	57.1	52.0	52.0	59.1	59.6
DeepSWE 1.1	79.0	78.8	67.4	74.1	74.2
SWE-Marathon	57.5 (xhigh)	53.1	45.6	42.5	46.3
ProgramBench	59.0	50.0	55.5	52.5	70.5
AA Intelligence Index v4.3	56	52	53	53	58

Three caveats matter. The Personal Superintelligence Index, MemoryBench and FocusBench are internal Meta evaluations. The model card says competitor runs are best-effort and may not be tuned. SWE-Marathon results for rivals used Codex and Claude Code, so that row is not a common-harness comparison.

The spread tells the story. Muse Wonder scores 85.6 on MemoryBench, the top result among all five models listed. On ProgramBench it scores 50.0 against 70.5 for Claude Opus 5.5 — a 20.5-point gap on the model's worst row. On Harvey LAB-AA, a legal benchmark, Muse Wonder Full's 13.1 triples Claude Opus 5.5's 4.2.

Coding is where Muse Wonder Full lands its strongest blows. It leads DeepSWE 1.1 at 79.0, ahead of Claude Opus 5.5 at 74.2 and GPT-6 Astra at 74.1. It leads SWE-Marathon at 57.5, an 11.2-point margin over Claude Opus 5.5. Claude Opus 5.5 still controls the general-coding and computer-use rows: ProgramBench at 70.5, OSWorld V2.1 at 81.8, Terminal-Bench 4.0-AA at 59.6 and the AA Intelligence Index v4.3 at 58.

How does Muse Wonder compare on price and specs?

Feature	Muse Wonder	Muse Wonder Full	Claude Opus 5.5	Claude Fable 5.1	GPT-6 Astra
Developer	Meta	Meta	Anthropic	Anthropic	OpenAI
Context	1M	1M	1M	1M	1.05M
Input modalities	Text, image, video, audio	Text, image, video, audio	Text, image	Text, image	Text, image
Open weights	No	No	No	No	No
API price (in/out per 1M)	$4 / $20	Subscription only	$4 / $20	$10 / $50	$10 / $50
Access	Model API, Muse Code	Muse Code Power ($50) or Maximum ($200)	Claude API	Claude API	OpenAI API
DeepSWE 1.1	78.8	79.0	74.2	67.4	74.1

Muse Wonder's $4/$20 pricing matches Claude Opus 5.5 exactly and undercuts GPT-6 Astra by 60 percent on input and output. Cached input for Muse Wonder costs $0.20 per million tokens. Muse Wonder Full has no API price at all: it exists only inside Muse Code subscriptions, at $50 for Power or $200 for Maximum per month.

The video and audio input support is the one spec-table difference no competitor matches at this tier. Both Anthropic models and GPT-6 Astra accept text and image only.

How do developers get started?

The setup path runs through dev.meta.ai and takes six steps, plus an optional terminal install:

1. Open dev.meta.ai. Click Start building to sign in. The Muse Code install command sits on the same page.
2. Pick your model. Muse Wonder costs $4 input, $0.20 cached and $20 output per million tokens. Muse Wonder Full requires a Muse Code plan: Power at $50 or Maximum at $200 per month.
3. Sign in with your Meta account. Confirm your name, accept the Model API terms, and click Get started.
4. Add billing, then create an API key. New accounts see a "No payment method" banner, and Create API key stays disabled until a card is added under Billing. Export the key as MODEL_API_KEY, per the quickstart.
5. Test it in the Playground. Open Chat, pick muse-wonder-1.0, and set temperature, max tokens and reasoning effort.
6. Copy working code. Click Get code for a curl or Python snippet. It calls the Responses endpoint at https://api.meta.ai/v1/responses.
7. Or use Muse Code in the terminal. On Mac, install with curl -fsSL https://dev.meta.ai/install.sh | bash. On Windows, use irm https://dev.meta.ai/install.ps1 | iex. Run muse in your project folder, trust the workspace, and sign in.

The API is OpenAI-compatible, so the OpenAI SDK works with a new base URL:

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
resp = client.responses.create(model="muse-wonder-1.0", input="List 3 risks in this lease clause: ...")
print(resp.output_text)

One detail trips up new Muse Code users: the CLI defaults to muse-spark-1.2, not Muse Wonder. Switch with /models, or launch with muse --model muse-wonder-1.0.

The OpenAI-compatible endpoint lowers switching costs to near zero for existing OpenAI SDK users — a code change of one base URL and one model name.

What is still missing?

Meta has not disclosed parameter counts for either model. Neither carries open weights, a break from the expectations Meta's earlier releases set. No full safety report accompanies the launch. And the strongest headline numbers — MemoryBench, the Personal Superintelligence Index — come from Meta's own internal evaluations, while competitor runs in the model card are best-effort and may not be tuned.

Independent verification will determine whether Muse Wonder Full's DeepSWE and SWE-Marathon leads hold under common-harness testing. Until then, buyers comparing it with Claude Opus 5.5 face a split verdict: Meta leads on software engineering and long-horizon memory benchmarks at half the modal price, Anthropic leads on general program synthesis and computer use.

Meta has also locked its top model behind its own subscription. Muse Wonder Full reaches developers only through Muse Code at $50 to $200 per month, which means Anthropic and OpenAI retain the only frontier-tier API offerings in this comparison.