Every few weeks a new model appears promising to change everything. The latest move comes from Meta with Muse Wonder and Muse Wonder Full, two models specifically geared towards long, autonomous work: programming, legal and financial tasks, and computer use. For a private company evaluating projects with AI agents, the question is not whether the model is impressive, but what changes —if anything changes at all— in its day-to-day operations. Let’s read the announcement with a business mindset, not a headline one.
(Source of the announcement: MarkTechPost, «Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It», October 2026.)
What Meta has launched, in two minutes
According to the published information, Muse Wonder is Meta’s most capable model to date and sits above its previous model, Muse Spark. The points that matter:
- 1 million token context and text, image, video and audio input; the output is text.
- Only on Meta’s cloud, via API or its command-line tool (macOS, Linux and Windows). There are no open weights or self-hosted installation.
- Two standout behaviours: active memory (it maintains the user’s context across conversations) and active focus (it persists with difficult tasks, tries alternatives and checks in before important actions).
- API pricing: 4 dollars per million input tokens and 20 per million output tokens. The Full version is only available within paid subscription plans.
How to read the benchmarks without getting carried away
The announcement includes a comparison table against models from Anthropic and OpenAI. The honest summary is that there is no absolute winner: the Full version leads 6 of 14 rows, stands out in memory tests and in software incident resolution, and clearly falls behind in others, such as ProgramBench, where Claude Opus 5.5 takes a wide lead.
There are three nuances that any manager should bear in mind before making decisions based on a table of this kind:
- Several tests are internal to the manufacturer. This is not necessarily bad faith, but an exam designed in-house rarely harms the host.
- The competition’s runs are "best effort". The material itself acknowledges that rival models may not be optimally tuned and that some comparisons do not use the same test environment.
- No benchmark measures your process. A model may shine at resolving code incidents and perform mediocrely at classifying your supplier invoices, which is what matters to you.
The practical conclusion: benchmarks are useful for ruling out, not for choosing. To choose, you need a test with your data and your use case.
What these launches mean for AI agents in an SME
The most relevant piece of information in the announcement is not a score, but where the market is heading. The major manufacturers are no longer competing to answer a question better, but to sustain multi-step tasks: planning, delegating to subprocesses, handling browsers and applications, and remembering context across sessions. This is exactly the territory of AI agents applied to real administrative processes.
This has two readings for a medium-sized company. The good one: capabilities that were experimental two years ago are starting to be available by subscription, without setting up infrastructure. The less comfortable one: the more autonomous a system is, the more governance, traceability and permissions matter. An assistant that drafts a document carries a limited risk; an agent that operates on your ERP, your email and your online banking does not.
Where it fits in real work today
Without straying from what we already do with our clients, there are areas where this generation of models delivers measurable value:
- Document reading and classification: delivery notes, rental contracts, insurance policies, specification documents from an industrial client.
- Preparing information for decision-making: summaries of cost deviations by project, by vehicle or by site, with the data coming from the ERP and not from intuition.
- Support for the helpdesk and back office teams: proposed responses, verified by a person before being sent.
- Intelligent automation: combining Power Automate with a model for the steps that previously required human judgement.
In all cases, the pattern we recommend is the same: the model proposes, a person validates and the system leaves a trace of what it has done.
Three questions before switching models
Before moving a project to a new model because it has been in the news, it is worth answering these:
Where does my data live? Muse Wonder runs exclusively on its manufacturer’s cloud. For a company with sensitive information —customer data, commercial terms, employment information— that is not a technical detail, it is a regulatory compliance decision.
How much will it really cost me? The price per million tokens is deceptive. What drives the bill is actual consumption: an agent working with a one-million-token context and retrying tasks can multiply the forecast expense. Always ask for an estimate based on real volume.
What happens if the provider changes the rules? Pricing, availability and conditions for these models change every few months. Designing the integration so that the model can be replaced without rebuilding the process is, today, a survival decision rather than one of technical elegance.
How we put it into practice at TISA
At TISA we have been supporting private companies with their management technology since 1987, and our approach to AI is the same as with an ERP: start with the problem, not with the tool. Our AI and data service begins by identifying what data you have and what state it is in, defining governance protocols aligned with regulations, evaluating tools according to your real investment capacity and developing a specific, measurable use case before scaling. Without organised data, no model —however well it scores— is going to perform.
We work with retail, distribution, construction, the food industry, hospitality, sports centres and fleets, and in all of them the AI that works is the one that relies on the management system you already use, usually Microsoft Dynamics 365 Business Central and Power Platform.
Do you want to know whether your company has an AI use case today that stands up? Write to us at info@grupotisa.com or call us on (+34) 971 305 885 and we will assess it with you with no obligation. You can see more about our services and cases at grupotisa.com.
