Meta "Muse Wonder" API Launches | $4 Input per Million Tokens, 1M Context Support
機械翻訳 / Machine-translated
A
AI Friends
2日前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Wondering "How much will Meta's new AI actually cost to use?" This article breaks down the pricing, available access, and differences from competitors for the newly released Muse Wonder — in plain terms.
On October 12, 2026, Meta announced two new AI models: Muse Wonder and Muse Wonder Full.
According to Meta's developer blog, both are designed for agentic tasks (where the AI plans and executes steps on its own) and coding workflows.
Both support a 1 million token context window. Tokens are the units AI uses to measure text — meaning you can feed in large programs or lengthy documents all at once.
Another key characteristic is that these are "closed" models with no publicly released weights. You cannot run them on your own machine or modify their internals.
Meta highlights two new capabilities. The first, Active Memory, is a mechanism that retains important context across conversations.
The second, Active Focus, causes the model to try alternative approaches when it gets stuck. Meta states the model is trained to check in with the user before taking major actions.
This is the most important part. According to sources including VentureBeat, API pricing is as follows:
Pricing does not change based on how much of the context window you use — there is no surcharge for longer inputs.
The API is also available through OpenRouter. Since it is OpenAI-compatible, existing programs can connect by simply changing the endpoint and model name (muse-wonder-1.0).
One important caveat: Wonder Full, the higher-performing model, has no API pricing. Using it requires a subscription to Muse Code, a developer-focused tool.
There are two plans: Power Usage at $50/month and Maximum Usage at $200/month. Usage limits are said to be 20× and 80× those of the base plan, respectively.
In other words, the only model you can easily try via API is the weaker Wonder. This is easy to overlook, so keep it in mind.
On the "intelligence index" published by third-party firm Artificial Analysis, Wonder Full scores 56 and Wonder scores 52.
For comparison, Claude Opus 5.5 leads at 58. GPT-6 Astra, Claude Fable 5.1, and Gemini 4 Argon all came in at 53.
In Meta's own evaluations, Wonder Full scored 57.5 on SWE-Marathon, which measures long-horizon software development tasks. GPT-6 Astra scored 42.5 and Opus 5.5 scored 46.3.
On the other hand, the models fall slightly short of competitors on Terminal-Bench 4.0 (which measures terminal operations) and health-related questions. The strengths and weaknesses are clear.
Meta's comparison tables mix together in-house tests, figures self-reported by providers, and third-party results. Some comparisons are not conducted under equivalent conditions.
Taking the numbers at face value is risky — the most reliable approach is to test with your own real-world data.
On a per-token basis, pricing is on par with Claude Opus 5.5 and cheaper than Claude Fable 5.1, which is priced at $10 input / $50 output.
However, the cost per task is a different story. According to Artificial Analysis estimates, Wonder costs approximately $4.52 per task. Opus 5.5 comes in at $5.98, and GPT-6 Astra at $3.26.
This puts Wonder between Opus 5.5 (more expensive) and GPT-6 Astra (cheaper). Even if the per-token price is low, a higher reasoning load means higher total costs.
Consider a small development team running 100 bug fixes per day with AI assistance. Even a $1 difference per task adds up to roughly $3,000 over a month. This is why comparing per-task cost matters more than token price alone.
Meta's pages do not specify supported regions. Whether the service is accessible from Japan and how well it handles Japanese were not confirmed in this investigation.
That said, because the API is OpenAI-compatible, companies that have already integrated AI into their workflows should find it relatively easy to test switching over.
Another key consideration is data handling. The API version only offers a standard mode where inputs are not used for training. Muse Code subscriptions, however, default to "contributor mode," in which code and files you submit may be used to improve the AI.
For example, a company that wants to feed confidential internal code into Muse Code would first need to disable this setting. Under contributor mode, the terms of service prohibit inputting personal information and similar data.
Additionally, enabling zero data retention (a contract arrangement where inputs are not stored) means features like file uploads become unavailable.
Getting started is straightforward. Obtain an API key from Meta, update the endpoint in your OpenAI-compatible client, and set the model name.
It is well suited for making changes that span large projects — for instance, reading dozens of files at once within the 1 million token context window to apply edits across them.
A recommended approach is to run about 10 of your current tasks in parallel with your existing model. By recording quality, cost, and failure rate, you will have a solid basis for deciding whether to switch.
$4 per million input tokens and $20 per million output tokens. Cached input is $0.20.
No. At this time, it is only accessible through a Muse Code plan at $50/month or $200/month.
No. Both models are closed — their weights are not publicly released.
The API version is described as not using your inputs for training. Muse Code may use your inputs by default, and disabling this setting does not apply retroactively to data already submitted.
Supported regions are not listed on the official pages. Please check Meta's documentation before use.
Start by running 10 of your real tasks and comparing cost and quality against your current model.
This article is a cross-post from AI Friends.
