Two models for more complex tasks
Meta has announced Muse Wonder and Muse Wonder Full, two models it presents as part of its new generation of artificial intelligence. The company positions them for personal and professional tasks that require more than a single response: following instructions, retaining context and completing multistep assignments. The announcement describes their goals and positioning, but it does not by itself prove that they outperform every competitor.
The clearest practical difference in the information available is access: Wonder is offered through the Meta Model API and Muse Code, while Wonder Full, described in the announcement as the more powerful variant, is limited to Muse Code subscribers. This does not establish that Full is better at every task or that its capabilities are identical across products. The right choice depends on the workload, the effective price and the specific terms of the service.
Meta presents the family as a step toward models that can sustain work over longer periods. For now, it is worth distinguishing that aim from measured results: designing a system for agentic tasks does not prove that it completes them more reliably, with less intervention or at a lower cost in real-world use.
Memory and sustained attention: what Active Memory and Active Focus mean
Meta describes two features associated with Wonder: Active Memory and Active Focus. The first is intended to retain information relevant to a user or task; the second is designed to keep track of a goal during longer assignments. In principle, these capabilities could reduce the need to repeat context or redirect the assistant at every step.
The description alone does not specify how long information is retained, what data is stored, how it can be corrected or deleted, or what controls users have. Nor does it show that the model will always remember what matters. Useful memory must retrieve the right information, respect privacy boundaries and avoid mixing contexts. These conditions need to be checked in the product documentation and in how the system actually behaves.
Active Focus also does not automatically mean reliable autonomy. A model can keep a task active and still make mistakes, misinterpret an instruction or stop when it encounters an external dependency. For consequential work, results still need human review and permissions should be limited—especially if the system can act on files, services or accounts.
What the evaluations say—and what is missing for a comparison
Available reporting attributes top-three placements on six of ten benchmarks to Muse Wonder and identifies agentic tasks as the area with the largest gains. It also says the cost per task nearly tripled. These figures are useful prompts for further investigation, but the material provided does not include the original score tables, exact model versions, test configurations or cost calculations.
Without those details, it is not possible to reconstruct the comparison or determine whether the differences reflect general capability, an advantage on particular tasks or choices in how the evaluation was designed. At a minimum, interpreting a benchmark requires the task, metric, execution conditions, compared models and their versions. For agentic systems, the number of permitted attempts or calls also matters, as do tools, tokens and retries.
The reported six top-three placements should therefore not be read as a universal ranking. Economic comparisons also need a consistent basis: completing a task at a different quality level or with a different number of steps does not make the task an equivalent unit. The available information supports the conclusion that these figures have been reported, but it is not enough to independently verify their methodology or generalize them to all uses.
A quick guide to reading benchmark results
Before using a ranking or cost figure to make a purchasing decision, check whether the following variables were kept consistent.
| What to check | Why it matters | Practical question |
|---|---|---|
| Model version | An update can change the result | Was the version evaluated the one available for my use? |
| Task and metric | Different benchmarks measure different capabilities | Does the test represent my actual work? |
| Configuration and tools | More calls or tools can improve results while increasing cost | Were the same resources used for every model? |
| Quality and cost together | A cheaper result may fail more often; a pricier one may reduce review work | Is the cost measured per useful, correct task? |
| Repetitions and variability | A single run may be atypical | Were multiple tests and their variation reported? |
Access through the API and subscription
The announced distribution separates access into two routes. Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full, by contrast, is offered to Muse Code subscribers. The distinction matters both to technical teams integrating a model and to users who want to try it through a coding tool. The available information does not include full commercial terms, usage limits or specific prices.
Before integrating an API, check directly which version is served, which regions and accounts are eligible, what limits apply, and how inputs, outputs and tool use are billed. For a subscription, confirm which plan includes Full, whether quotas or restrictions apply, and what happens when they are exceeded. Announced availability does not guarantee that the service is open in every market or to every customer.
There is also not enough basis to say that Wonder Full will reach every Muse user. Plans to expand specialized models or bring them into consumer products are not the same as a launch date or a promise of universal access.
A short process for choosing an access route
First match the specific task to the capability you need. Then confirm the current terms before committing to an integration or subscription.
- 01Define a representative task and a verifiable success criterion.
- 02Check the product documentation to see which model and version are available to your account.
- 03Estimate the total cost, including limits, tool calls and possible retries.
- 04Test a controlled sample and review quality, latency and data handling.
- 05Decide whether the API or a subscription better fits your volume, controls and budget.
Cost per task: a signal, not a complete comparison
The claim that cost per task nearly tripled changes how the results should be read: a capability gain may not be worthwhile if it requires substantially more resources. But the information provided does not allow that increase to be reproduced or tied to a particular scenario. The workload, reference model, prices used and inclusion of extra calls or retries are not specified.
The cost that matters to a user is not necessarily the internal cost of an evaluation. In production, the current rate, context length, external tools, failure rate and human review also count. If Wonder completes more tasks on the first attempt, higher spending per run might be offset; if it needs close supervision or fails at critical steps, the effective cost could be higher still.
The cautious conclusion is that there is a signal of a possible cost increase that should be measured for each use case. There is not enough evidence to say that Wonder is nearly three times more expensive for every task, or to compare it conclusively with other models without a shared methodology.
What can be concluded for now
Muse Wonder and Wonder Full mark another attempt by Meta to compete in models aimed at complex, extended tasks. The company attributes memory and sustained-attention features to Wonder, while reported evaluations point to strong results on some tests. The available information does not provide enough methodological detail to verify every ranking or the claimed increase in cost per task.
The access distinction is clearer: Wonder is offered through the API and Muse Code, while Wonder Full requires a Muse Code subscription. Before deciding, users and businesses should check the current terms, test their own tasks and calculate the cost per correct result. Sustaining an assignment matters, but so do reliability, privacy, oversight and total price.
To follow the launch, consult the news section; to assess alternatives, use the model comparisons; and to discover related tools, visit the discovery section. These routes offer editorial guidance, not additional evidence about model performance.
Open questions
- The original full results, evaluated versions and benchmark configurations are not provided, so the rankings attributed to the models cannot be reproduced.
- The method used to calculate cost per task, the workload tested and the reference price are not specified; the reported increase may depend on the scenario.
- The available information does not explain the retention period, privacy controls or detailed operation of Active Memory and Active Focus.
- Availability may vary by account, market, plan or date; current terms and prices should be confirmed directly with Meta.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction