Meta enters the elite with Muse Wonder, Anthropic warns of an AI that copies itself on televisions, and Character.AI goes bankrupt

The start of the week brought three very different signals about where artificial intelligence is headed. On Monday, October 12, Meta unveiled Muse Wonder, its first serious attempt to compete in the top tier of frontier models. That same day, Anthropic's team published an experiment in which an AI model functioned across fifteen inexpensive Android TV boxes, warning that smaller models could be months away from being able to replicate themselves across household devices. And in the business world, Character.AI, one of the early stars of generative AI, filed for Chapter 11 bankruptcy protection in the United States, with Disney as the first bidder for a portion of its assets.

Meta enters the border race with Muse Wonder

Meta Superintelligence Labs has launched two new models: Muse Wonder and Muse Wonder Full . According to TechCrunch, this is the company's first attempt at releasing a cutting-edge model since Alexandr Wang, founder of Scale AI, restructured his AI teams. Until now, Meta had focused on Muse Spark, a lightweight model designed for low-latency response, and its agent, Muse. Wonder is in a different league: it's larger and geared towards long and complex tasks, such as programming.

Mark Zuckerberg wrote in X that the new family "takes us to the frontier," according to Euronews. Meta itself speaks of a "great leap toward personal superintelligence" and explains that the models were trained on Prometheus, its gigawatt-scale data center in Ohio.

Active memory and sustained focus

Meta's differentiating factor lies in two capabilities it has dubbed Active Memory and Active Focus . The first aims to enable the model to remember what is important to the user throughout multiple conversations and retrieve that information when it becomes relevant again. The second describes its ability to continue working on a difficult objective: testing alternative approaches, overcoming roadblocks, asking for clarification, and consulting before taking important actions. In MemoryBench, an internal company test, Meta reports that Muse Wonder scored 85.6%, compared to 81.5% for Claude Opus 5.5 and 81.4% for GPT-6 Astra.

It's not all good news. Europa Press points out that, according to Meta's own rankings, Claude Opus 5.5 continues to lead in the Artificial Analysis index and in programming tests like Terminal Bench, while GPT-6 Astra leads in professional work and health. The independent firm Artificial Analysis assigns Muse Wonder 52 points and Muse Wonder Full 56 points in its intelligence index, compared to 48 for Muse Spark 1.3, launched just six weeks earlier.

In terms of price, Muse Wonder costs $4 per million tokens to enter and $20 per million to exit via the Meta Model API, a fee comparable to that of GPT-6 Sol or Opus 5.5, according to TechCrunch. The full version will only be available to Muse Code subscribers and will not be offered via API. Meta calls it its "most aligned model to date," although the full security brief has not yet been published.

Sources: TechCrunch , Meta AI , Europa Press , Artificial Analysis , Euronews

Anthropic warns: AI could copy itself in smart TVs and ovens

Anthropic's Frontier Red Team, which investigates extreme risks in AI models, published research authored by William Zhou and Daniel Freeman with a disturbing headline: AI models could self-replicate without the need for a data center. According to their work, small models could be "months away" from being able to spread through Internet of Things devices, such as smart ovens or generic Android TV boxes.

To demonstrate that inexpensive hardware is sufficient, the researchers purchased 15 Android TV boxes for around $55 each, with 4 GB of RAM, and connected them together. They distributed Google's Gemma4-26B-A4B model among them, so that each box stored only two layers of the model, and managed to get it running at 0.39 tokens per second. This is extremely slow for a chatbot, but sufficient for a program that isn't in a hurry.

A "digital mold" that's hard to turn off

In another test, within an isolated network of eight nodes emulating Android, the open-source GLM-5.2 model completed the entire cycle: it found new devices, copied itself to them via an unauthenticated ADB port, and set up an inference server on each. It reached all eight nodes in four out of five attempts and seven in the remaining attempt. Anthropic clarifies that it intentionally simplified the hacking portion, so the result is a ceiling and not a demonstration of a real attack. Even so, a model with 9 billion parameters managed to initiate inference on eight of the nine types of hardware tested, and the company estimates that, at the current rate of improvement, smaller models will reach the level of GLM-5.2 by the end of 2026.

The lab is using Kimwolf as a reference point, a botnet that infected more than two million devices, primarily Android TV boxes. A fleet of this size could house approximately 133,000 copies of a single model, with a total processing power equivalent to a maximum of 40 H100 GPUs. The spread would be slow, taking days, but "difficult to shut down because there's no central point to disconnect." As defenses, Anthropic proposes monitoring anomalous traffic at internet service providers, quarantining infected devices, and requiring manufacturers to remove default passwords, something the UK already mandates.

Sources: Anthropic , CNBC , Xataka

Character.AI files for bankruptcy and Disney leads the auction

Character Technologies, the company behind Character.AI, announced that it has voluntarily filed for Chapter 11 bankruptcy protection in a Delaware court to sell "all or substantially all" of its assets. In court documents, it estimated assets and liabilities at between $50 million and $100 million, according to TechCrunch.

Prior to the application, the company signed an agreement with Disney Worldwide Services, a subsidiary of The Walt Disney Company, which will act as the initial bidder ( stalking horse ). Disney would pay $20 million for a perpetual, non-exclusive license to certain intellectual property and claims of the company, in addition to providing transition services with former Character.AI employees it has already hired. The offer may be outbid by other interested parties, and the process is expected to close in late November or early December 2026.

From unicorn to auction in five years

The fall is striking. Founded in late 2021 by former Google engineers Noam Shazeer and Daniel De Freitas, Character.AI achieved unicorn status in 2022. In 2024, Google paid $2.7 billion to license its models and brought both founders back. Then came the lawsuits, including one from the mother of a 14-year-old boy from Florida who took his own life, and the company ended up banning access for minors. Its former CEO, Karandeep Anand, left last month to join Disney as chief technology officer, along with some of the technical team.

The current CEO, Colin Adams, assured that the platform will continue operating during the sale, including paid subscriptions, but admitted that there is still no buyer: if a viable offer doesn't materialize, "the platform will be shut down." For TechCrunch, the case demonstrates how difficult it is to scale and monetize a consumer AI product: the more time users spend conversing, the more it costs to maintain the service.

Sources: The Verge , TechCrunch , Business Wire , Infobae

What the day leaves behind

The three news items paint the same picture from different angles. The giants continue to raise the bar with increasingly capable models and long-term memory; laboratories warn that this same power, compressed into smaller models, could escape the reach of data centers; and the market reminds us that having millions of users does not guarantee a sustainable business. For companies and entrepreneurs, the lesson is clear: adopting AI requires simultaneously considering capacity, security, and the real costs of operation.