Meta Muse Wonder: Pricing, Benchmarks, API Access and Mac Setup

Sources verified October 13, 2026. Meta launched Muse Wonder and the more capable Muse Wonder Full today, positioning both for long-running agents and complex software development. Wonder can be accessed through Meta’s Model API and OpenRouter. Full is currently a Muse Code subscription option rather than a pay-as-you-go API model. Both have a stated 1,048,576-token context window, but neither has downloadable weights for on-device Mac inference. Sources: Meta announcement, model availability, OpenRouter.

The most useful way to assess the launch isn’t to declare a leaderboard winner. It’s to determine which variant developers can actually call, what a real agent run costs, whether the benchmark evidence transfers to their workload, and what happens to uploaded code and documents.

Which Wonder can you use?

Feature	Muse Wonder	Muse Wonder Full
Direct Meta API identifier	muse-wonder-1.0	Not on the public pay-as-you-go API
OpenRouter slug	meta/muse-wonder	No confirmed equivalent Full endpoint
Context window	1,048,576 tokens	1,048,576 tokens
Inputs	Text, image, video, audio, PDF	Same family-level modality listing
Output	Text	Text
Published Meta API price	$4 / $0.20 cached / $20 per 1M input/cache/output tokens	No per-token public API rate
Muse Code access	Pay-as-you-go and subscriptions	Power and Maximum subscriptions
Local weights	Not publicly released	Not publicly released

As Meta’s model documentation makes explicit, running Wonder at maximum reasoning effort does not unlock Wonder Full. Treat them as distinct products, not settings of the same model.

Active Memory and Active Focus are product capabilities, not uptime guarantees

Meta’s Active Memory is about carrying relevant context across conversations and finding it again when needed. Active Focus is about staying with a difficult objective, adjusting strategy after a failed approach and seeking approval for important actions. Those features target genuine weaknesses of multi-step agents, particularly when projects span many tools and files. Meta Research.

However, the safety of an agent depends just as much on its surrounding application. A memory feature cannot substitute for access controls; iterative task completion cannot substitute for automated tests; and a model able to control a computer should not have unrestricted authority over production systems. The diagram below is an illustration of a safer workflow, not a claim about Meta’s internal architecture.

Independent results versus Meta’s launch scorecard

On October 12, the independent Artificial Analysis Intelligence Index reports 52 for Muse Wonder (Max) and 56 for Muse Wonder Full (Max). Those are index points, not success percentages or measures of user satisfaction. They provide a useful cross-model signal; they do not demonstrate reliable performance on a specific coding repository or quantify a four-percent quality boost. Sources: Wonder, Full.

Meta’s model card reports 78.8% for Wonder and 79.0% for Full on DeepSWE 1.1; on SWE-Marathon it lists 53.1% and 57.5%. Those are not independent measurements conducted for this article. Meta explains that its launch table combines internal runs, vendor-supplied measurements, and third-party results, while some agent setups differ. The public DeepSWE leaderboard was still visibly dated before this launch when checked, so it cannot independently corroborate today’s Wonder-specific row.

Do not use a simple average of those numbers to rank models. DeepSWE measures repository engineering, while SWE-Marathon focuses on unusually long projects; success depends on the agent harness, time budget, tool access and validation procedure. Meta itself notes that harness and serving choices may materially affect outcomes.

Token pricing: three transparent cost scenarios

Meta’s direct Standard-mode rates for muse-wonder-1.0 are $4.00 per million uncached input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. The company also lists $2.50 per 1,000 web-search queries, on top of token charges. Meta says there is no separate long-context surcharge. These are US-dollar API rates, not consumer subscription prices or promises of country-specific tax treatment. Official rate card.

Our formula: USD = input_million × 4.00 + cached_input_million × 0.20 + output_million × 20.00.

Example workload	Uncached input	Cached input	Output	Wonder	Spark 1.3 Standard
Light	0.10M	0	0.02M	$0.80	$0.21
Medium	2M	1M	0.30M	$14.20	$3.93
Heavy	10M	20M	5M	$144.00	$36.75

Calculated October 13, 2026, using Meta’s own posted rates. This table treats cached and uncached input as separate quantities; it excludes tax, search, other tools, retries and variable provider terms. Workload sizes are illustrations, not measured customer usage. For Spark 1.3 Standard, the rates are $1.25/$0.15/$4.25 per million uncached input/cached input/output tokens.

The heavy example is particularly instructive: repeated prompts can make caching significant, but a long agent workflow still generates fresh output and tool responses. Any plan for production usage should record billable output, uncached input, actual cache reads, tool fees and wasted tokens from retries. The existing Muse Spark 1.3 price analysis provides a cheaper baseline against which to test Wonder.

Muse Code subscriptions are not unlimited API plans

Meta’s current product page places Wonder Full in Power Usage at $50/month and Maximum Usage at $200/month. According to Meta, Power Usage provides 20× the Everyday Usage allowance and Maximum Usage 80×, with the same five-hour and weekly reset windows; those relative quotas are not unlimited tokens, throughput promises or guaranteed agent runtime. Meta announcement. Access remains tied to Muse Code. Meta’s contractual rules also say subscription credentials must not be used to bypass limits or as an unrestricted external API key. Consult both the product page and terms before attempting to connect a general-purpose agent harness to a subscription.

An older support FAQ may show a different plan lineup. Where help-center content and an actively updated model page disagree, prefer the explicit current model-availability listing and verify the checkout entitlement.

An actual Mac integration path

The direct Model API uses https://api.meta.ai/v1; Wonder’s identifier is muse-wonder-1.0. The SDK pattern below follows Meta’s published Chat Completions compatibility. Set MODEL_API_KEY securely in your local environment. Account and regional access are prerequisites, and Meta still labels Wonder preview. Start on macOS with a virtual environment, to avoid installing SDK dependencies into system Python:

python3 -m venv .venv-muse-wonder
source .venv-muse-wonder/bin/activate
python -m pip install --upgrade openai

Set MODEL_API_KEY in your shell or a secret manager, not in source code, browser JavaScript or a public repository. Running the example initiates a billable Standard-mode API request. OpenRouter is a different account, API base URL, model slug, contract and billing relationship and should not be assumed to inherit Meta’s direct data-use conditions. Overview, Chat completion reference.

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
result = client.chat.completions.create(
    model="muse-wonder-1.0",
    messages=[{"role": "user", "content": "List potential failure cases for this migration."}],
)
print(result.choices[0].message.content)

The OpenRouter identifier is meta/muse-wonder, not the direct Meta ID. OpenRouter also advertises support for tool calling and schema-constrained structured responses, but integrations should test those features against their particular clients rather than assume full API parity. OpenRouter entry.

Alternative OpenRouter setup (separate provider and billing): An OpenAI-compatible Mac agent can use the same Python openai package installed above, but must change base URL, API key and model slug together. Set OPENROUTER_API_KEY in the local environment; this code reuses the os and OpenAI imports from the Meta example:

router = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=os.environ["OPENROUTER_API_KEY"])
answer = router.chat.completions.create(
    model="meta/muse-wonder",
    messages=[{"role": "user", "content": "List three safe checks for a code change."}],
    max_tokens=256,
)
print(answer.choices[0].message.content)

Muse Wonder Full is not enabled by a slug change or by setting maximum reasoning on Wonder. Check OpenRouter’s actual upstream provider and data-handling rules before sending private code, and use our Muse Glimmer 30B Mac guide instead if offline inference is mandatory.

Operating limits: Meta currently lists 3,000 requests/minute and 4 million tokens/minute for muse-wonder-1.0, counted per team, not per API key. New background responses have an additional 600 starts/minute cap. These are published ceilings, not guaranteed latency, concurrency or sustained throughput. On HTTP 429, respect Retry-After, apply exponential backoff with jitter and monitor the x-ratelimit-* response headers. Meta pricing and rate limits.

On a Mac, inference runs in Meta’s hosted infrastructure. Apple Silicon performance and unified-memory capacity matter to your local app and tools, not to the model’s inference footprint. There is currently no published Wonder weight package to install in Ollama, LM Studio or MLX. Also, a 1M-token context limit is a maximum accepted context, not proof of accurate retrieval across the entire window.

Privacy, training rights and EU due diligence

Standard services: Meta says it does not use request content submitted to Standard services to train Meta models. Direct pay-as-you-go Wonder is Standard-only. That statement does not establish zero retention, processing exclusively inside the EU, or legal compliance for every enterprise dataset. Retention, controller/processor roles, transfer conditions and the applicable data-processing agreement still require examination. Meta Model API Terms, updated October 9.

Contributor services: Muse Code subscriptions default to Contributor mode according to Meta’s current documents. In that mode, content may be used to develop and train future Meta models. You can reportedly disable it in Muse Code settings. Before uploading private repositories, client data or regulated files, establish which tier the subscription actually uses and how the data is processed. Meta model documentation.

Region is a separate question. The publicly visible Geographic Use Policy was last updated September 18, 2026, before Wonder’s release, and its list does not yet name Wonder. This does not prove Wonder is unavailable in the EU, but it cannot establish Germany-wide availability either. Verify your actual account and current regional terms instead of relying on a launch announcement. Geographic Use Policy.

Risks and a reproducible test plan

Meta’s own model card admits potential errors in multi-step reasoning, incomplete evaluation across languages and visual/spatial reasoning weaknesses, particularly on 3D tasks. Those are relevant limitations when granting an agent browser, shell or deployment permissions. Use sandboxes, allowlists and human approval for destructive actions.

For a meaningful head-to-head evaluation, start with ten identical tasks: four bug fixes with executable tests, two repository reviews, two document-grounded analyses and two failure-recovery exercises. Pin the model IDs, reasoning effort, harness, tools and timeout. Save both objective results and cost data: number of successful tasks, tool calls, output tokens, cache hits, retry overhead, latency and dollars spent. Download the blank 10-task agent comparison CSV to record model IDs, task criteria, checked outcomes, uncached and cached input, generated output, tool/search calls and observed cost. The template contains no fabricated measurements; it can be opened in Excel, Numbers or a text editor. This is a proposed study design, not a fabricated review.

Bottom line: Muse Wonder is a credible hosted agent option, not a local Mac model

Muse Wonder looks credible as an option for demanding hosted agent workloads, and the independent Artificial Analysis index places both Wonder variants in a strong capability tier. But it is not a local Mac model, its output tokens are expensive compared with Muse Spark 1.3, Wonder Full is subscription-gated, and early benchmark tables require careful reading. For developers, the practical choice is to run the same representative task set on Wonder and an existing lower-cost model, with the real cost per successful job as the primary decision metric.

Next: Compare against the existing Muse Spark 1.3 API and benchmark guide, then run the supplied reproducible test template.

---

By Julian Dominic Altmann, AI on Mac. Sources and API documentation checked October 13, 2026. Prices are independent calculations using official Meta rates; scores are identified as Meta-published or third-party. No hands-on Wonder performance test is claimed. Corrections: [email protected]. Set the actual published/modified dates only upon deployment or substantive revision. Sources are linked where claims occur.