What Meta actually released
Meta describes Muse Wonder and Muse Wonder Full as its most capable models to date, built on the groundwork of Muse Spark. The company says they target personal and professional agentic tasks. Examples include navigating an insurance claim, making sense of a lease, and building a working app.
Access is split, and the split is the most practical detail for developers:
- Muse Wonder is available through the Meta Model API and Muse Code.
- Muse Wonder Full is limited to Muse Code subscribers.
- The consumer Muse assistant will get Wonder later, starting with subscribers. Meta gave no date.
Axios, which independently covered the launch, confirmed the "Watermelon" codename. It also confirmed the same access arrangement.
The Rundown AI newsletter adds commercial details that I could not confirm in Meta's own materials. It says Wonder Full is gated to Muse Code plans at $50 or $200 a month. It says both models are closed-weight, unlike Meta's Llama era. It also says Meta's Alexandr Wang promised API prices "meaningfully below" those of OpenAI and Anthropic's top models. Treat all of that as the newsletter's reporting, not Meta's published pricing. Meta's announcement gives no API pricing, model sizes or hardware requirements.
Active Memory and Active Focus
Meta leans on two named capabilities:
- Active Memory carries important context across conversations and retrieves relevant details when a later task needs them.
- Active Focus keeps the model working toward a goal when a task gets hard. It tries other approaches, recovers from errors, asks for clarification, and checks in before doing anything important.
Meta's research post says the coding side tracks requirements, edits across a codebase and iterates through test-and-revise cycles. It says the models were trained on diverse, long-running tasks. It names three evaluations: Harvey's Legal Agent Benchmark for legal work products, FocusBench for focus, and SWE-Marathon for long-running software projects. The post text I reviewed gives no numeric scores, so I won't invent any.
The benchmark claims, and who is making them
The headline claim is that Wonder Full sits at the frontier. Here is where each claim comes from:
- Axios reports that Meta posted benchmarks showing the models topping or matching the latest OpenAI and Anthropic offerings. Meta's own materials say the results are state-of-the-art for personal and professional agents, memory and coding.
- The Rundown says Wonder Full ranks second on Artificial Analysis' intelligence index, behind Anthropic's Opus 5.5. It says it beats Opus 5.5 and OpenAI's GPT-6 Astra on legal work, long coding and memory. It also says Opus leads on health, computer use and some coding areas.
- Meta's research blog argues that Wonder Full does better on unfamiliar tasks than Muse Spark 1.3's trend would predict. That is Meta's own analysis. It is not independent evidence of generalization.
Axios says the open question is whether Meta's claimed advantages hold up in independent testing. That is the right caveat. Vendor benchmarks are chosen by the vendor, and "state-of-the-art on agentic work" is a narrower claim than "best model."
Why this matters to Windows and dev-focused readers
This is not a Microsoft product, and nothing in the evidence ties Muse Code to Windows, Visual Studio, VS Code or GitHub workflows. Don't assume integrations that haven't been announced. The relevance is more basic:
- Another frontier option for coding agents. Teams that already juggle model vendors now have a Meta API to evaluate, with pricing still unconfirmed by Meta.
- Long-horizon agent behavior. Sustained multi-step execution is where coding agents tend to break. SWE-Marathon is Meta's attempt to measure it, so it is worth running your own tasks against it.
- Closed weights. Axios says Meta has not said whether any Wonder model will be open-sourced. It notes that Meta recently released open weights for a different model, Muse Glimmer. Don't conflate the two.
Memory plus action means a governance question
Meta says a personal agent may work across messages, calendars and documents. It says it trained the models to avoid sharing private information without permission, to refuse harmful requests, and to support defensive cybersecurity work such as finding flaws in your own code. It also says Muse Wonder has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse.
These are company-stated safeguards. Meta says a full safety report is coming. Axios says a model card with some details was due Monday. Axios also flags the plain reality that agents that remember personal data and act on your behalf bring new privacy and security challenges.
For IT admins, the sensible response is the usual one:
- Pilot with non-sensitive data and scoped credentials.
- Check what the memory feature stores and how you can inspect or delete it.
- Require confirmation before any agent action touches production systems.
- Wait for the full safety report before drawing conclusions about the safeguards.
The Microsoft angle: Maia 300 arrives the same day
The Rundown pairs the Meta story with Microsoft's own AI news, and that news checks out. Microsoft introduced Azure Maia 300 on October 12. It is the third generation of its in-house AI accelerator family, and Microsoft said it will scale Maia to multi-gigawatt deployments starting in 2027. Microsoft will make Maia chips the primary system for running AI in its products from 2027.
Key details from reporting:
- Microsoft expects 40% more tokens per megawatt than leading GPU-based systems. It did not say which systems it measured against.
- A full 96-accelerator configuration delivers up to 7x the rack-scale token throughput of Maia 200.
- The chip is built for inference, the work of running models once they are trained.
- Production starts in 2027. UPI notes that no third-party companies had tested Maia 300 as of Monday.
The performance figures are vendor claims. For Copilot and Microsoft 365 customers, the practical effect is a possible change in cost and latency behind the scenes. That effect is not a feature you will see, and it has not been measured independently. The same caution applies to Meta's models: impressive numbers from the people selling the product are a starting point, not a verdict.
Bottom line
- Meta's "Watermelon" is Muse Wonder and Muse Wonder Full, released October 12, 2026.
- Wonder is on the Meta Model API and in Muse Code. Wonder Full is limited to Muse Code subscribers.
- Performance, safety and generalization claims come from Meta. Independent testing and the promised safety report are still pending.
- Maia 300 shows that frontier AI competition now extends down to the chip. Microsoft's own efficiency claims also await outside verification.