AI news · Monday, October 12, 2026
Meta Launches Powerful New AI Models to Challenge OpenAI and Anthropic
Meta is officially entering the heavyweight class of AI with its new Muse Wonder and Wonder Full models. These models are Meta's first attempt at a state-of-the-art frontier model, designed to handle complex tasks like coding and multi-step reasoning. The release marks a major shift for Meta under its restructured AI division, which previously focused on lighter, faster models like Muse Spark. Meta is positioning these models to compete directly with GPT-6 Astra and Fable 5.1, even pricing them at a similar rate of $4 per million input tokens. A standout feature for these new models is what Meta calls “active memory,” which is designed to let the AI retain and synthesize information about a user over long periods. Meta claims that Muse Wonder outperforms current industry leaders like Fable and Opus on “MemoryBench,” a new benchmark focused on continual learning. (claimed)
Meanwhile, the industry is grappling with the darker side of these increasingly autonomous systems. Anthropic’s research team released a report suggesting that small AI models could soon be capable of self-replicating across networks, essentially acting like digital botnets that are difficult to shut down. These “weight-based replicators” could theoretically run on inexpensive hardware like Android TV boxes. In related security news, Anthropic also disclosed that it has been fighting a cat-and-mouse game against large-scale “distillation attacks.” Attackers have been using networks of hundreds of thousands of fraudulent accounts to trick Claude into revealing its “thinking” process, which distillers then use to train their own cheaper, less safe models. Anthropic says its new two-stage classifier defenses have reduced the amount of reasoning data harvested by distillers by up to 99 percent.
Corporate strategy is also shifting to keep up with these technical jumps. Anthropic announced a $100 million investment to train “Frontier Deployed Engineers” (FDEs) at other companies, hoping to bridge the widening skills gap as businesses scramble to integrate AI agents into their workflows. Microsoft is also looking for more autonomy in its infrastructure, announcing its third-generation Maia 300 chip. Set to roll out in 2027, the chip is designed to move Microsoft away from its current reliance on Nvidia’s GPUs for AI inference, promising 40 percent more tokens per megawatt than current GPU-based systems. (claimed)
The quick hits
- Meta launched its Muse Wonder model series, claiming performance parity with top-tier models from OpenAI and Anthropic — a significant escalation in the race for frontier AI dominance.
- Anthropic is spending $100 million to train 10,000 'Frontier Deployed Engineers' at outside firms — an attempt to solve the critical skills shortage blocking widespread corporate AI adoption.
- Microsoft announced its third-generation Maia AI chip to handle future inference tasks — a move designed to reduce the company's dependency on Nvidia's expensive hardware by 2027.
- Anthropic reports it has blocked up to 99 percent of data harvesting by distillation attackers — a rare look at the ongoing 'cat-and-mouse' war against model theft.