Meta is back among the top tier of AI models. On October 12, its AI unit Meta Superintelligence Labs introduced Muse Wonder and the more powerful Muse Wonder Full, built for coding, long-running agent tasks, and operating computers. An independent evaluation confirms that Meta is moving close to OpenAI and Anthropic, though not to the very top. For readers in Germany, the more pressing question is a practical one: who can actually use these models, and on what terms?
Key takeaways
- Muse Wonder Full scores 56 points on the independent Artificial Analysis Intelligence Index, ranking fourth, tied with Claude Sonnet 5.5 and just behind Claude Opus 5.5 at 58.
- The standard Muse Wonder model costs $4 per million input tokens and $20 per million output tokens through Meta’s Model API, with a context window of one million tokens.
- The stronger Full version is not available through the API, only through a subscription to the coding tool Muse Code for $50 or $200 a month.
- With the subscription, inputs may be used for training by default; through the API they are not. The weights of both models remain closed.
- Meta has not published a country list for the API, and its consumer agent Muse is not yet available in Germany.
What Muse Wonder is meant to do
On its developer site, Meta describes Muse Wonder as a frontier model for coding, long-running agents, and computer use. It is designed to plan long tasks, hand off parts of the work to sub-agents, and recover from errors. According to VentureBeat, Mark Zuckerberg called the new models a much larger and more advanced class than the earlier Muse Spark family. Meta’s chief AI officer Alexandr Wang told Bloomberg that the model keeps pace with ChatGPT and Claude on coding as well as on financial and legal tasks.
For Meta, the launch marks a turning point. Some 14 months after reorganizing its AI division, the company had recently suffered setbacks, including the image model Muse Image, which Meta pulled three days after launch over controversial default settings, according to The Next Web. With Muse Wonder, the rebuilt unit is delivering a model that competes with today’s leading models in the rankings.
What the numbers show
Meta published its own comparison table, which does what every vendor table does: it shows the company’s model ahead where it shines. By that account, Muse Wonder Full leads on the long-horizon coding test SWE-Marathon with 57.5 points, ahead of Claude Opus 5.5 at 46.3 and GPT-6 Astra at 42.5. On Terminal-Bench, a test of command-line work, and on the medical benchmark HealthBench Pro, Opus 5.5 comes out ahead. VentureBeat notes that the table mixes results from independent evaluators with scores Meta measured itself, which makes comparisons harder.
The Artificial Analysis Intelligence Index, which combines many tests under identical conditions, is more reliable. There, Muse Wonder Full scores 56 points and the standard model 52. That puts the Full version level with Claude Sonnet 5.5 and ahead of GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 at 53 points each; Claude Opus 5.5 stays in front with 58. The cost per task is also worth a look: according to Artificial Analysis estimates cited by VentureBeat, an average index task costs $4.52 with Muse Wonder, $5.98 with Opus 5.5, and $3.26 with GPT-6 Astra. So Muse Wonder is no bargain option, but it is not the most expensive model either.
Who should take a look, and where the catch is
Muse Wonder is most interesting for developers building coding agents or long automated workflows. The model is available through Meta’s Model API under the ID muse-wonder-1.0. Cached input costs only $0.20 per million tokens, which pays off for agents that read the same code context again and again. According to VentureBeat, the limits are 3,000 requests and four million tokens per minute, and web search costs $2.50 per 1,000 queries.
Anyone who wants the stronger Full version cannot get it through the API, only through a subscription to Muse Code, Meta’s coding tool, which we covered at its launch in August (in German). The Power Usage tier at $50 and the Maximum Usage tier at $200 a month offer 20 and 80 times the base allowance, respectively. The difference in data handling matters: through the API, Meta promises not to use submitted content for training. Muse Code subscriptions, by contrast, default to the so-called Contributor mode, which allows exactly that. Anyone working with company code there should check the setting before getting started.
The biggest open question for German users is availability. Meta publishes no country list for the Model API, and none of the announcements explicitly mention Europe. Meta’s consumer agent Muse is not yet available in Germany, and there is no launch date for the EU. Developers here should check directly with Meta whether they can use the API and Muse Code without workarounds before making any plans.
Open weights? Not this time
With its earlier Llama models, Meta was the most important provider of freely downloadable model weights. Muse Wonder breaks with that approach: the weights of both versions are not public. According to The Next Web, Meta has released a scaled-down model with 30 billion parameters as open weights, but it is keeping the flagship to itself. That also has regulatory consequences. Under the EU AI Act, truly openly licensed models are exempt from certain documentation duties as long as they are not classified as posing systemic risk; Meta is not taking that route for Muse Wonder. Meta has not signed the EU’s voluntary code of practice for providers of large AI models.
In its technical documentation, the so-called model card, Meta addresses risks in cybersecurity, biological and chemical hazards, and autonomous behavior by the models. For consequential agent actions, Meta recommends human confirmation and acknowledges that the models still make mistakes.
Bottom line: more competition at the top, but not for everyone
Muse Wonder is Meta’s most convincing answer yet to OpenAI, Anthropic, and Google. Fourth place on the independent index, strong results on long coding tasks, and pricing in between its rivals make the model a real alternative, and more competition at the top tends to push down prices elsewhere too. For readers in Europe, the benefit remains limited for now: as long as Meta gives no clear answer on availability and offers its strongest model only through a subscription that allows training by default, Muse Wonder is a model to watch rather than one to switch to right away. The interesting question is whether Meta can hold its lead on long agent tasks once the competition ships its next models.