Meta: Muse Wonder has arrived — by the company’s own account, its first frontier model family. On Monday, Meta Superintelligence Labs unveiled Muse Wonder and Muse Wonder Full, models designed for personal and professional agent tasks; according to Meta, Active Memory and Active Focus support long, multi-day work. Wonder is available through the Meta Model API and in Muse Code (according to VentureBeat, about $4 per 1M input and $20 per 1M output tokens), while the more powerful Wonder Full is for now limited to paying Muse Code subscribers. The models are also coming to the Muse agent, first for subscribers. Independent rankings (e.g., Artificial Analysis) have put Wonder Full at the top of the field; the lack of open weights and the API restriction on Full are being debated. Source: Meta Research, TechCrunch.
Anthropic: small models could be able to self-replicate on IoT devices within months. According to Monday research from the Frontier Red Team, small open-weight models are already close to being able to replicate on commercial IoT devices (e.g., smart TV boxes, smart ovens) without a data center. The team ran Gemma4-26B-A4B on 15 Android TV boxes costing about $55 each; in the simplified, firewalled test, GLM-5.2 was the only open-weight model examined that reliably completed the full replication chain. The authors stress that the cyber-intrusion step was deliberately made easy, so the result is an upper-bound estimate of real-world success. Source: Anthropic Research.
Microsoft: Maia 300 will become the main inference platform for its own products — multi-gigawatt deployment from 2027. Scott Guthrie, head of the cloud and AI business, announced the Maia 300 system and said that from 2027, with multi-gigawatt Maia deployments, Maia will become the primary inference system for Microsoft’s own products and services. The company claims Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU systems, and up to seven times the token throughput of Maia 200 at rack level. Maia 200 is already running in production (including on Copilot workloads). Source: Microsoft Blog, The Verge.
Epic Games: “AI-first” development on the Disney shooter — Sweeney confirmed the shift. According to a Bloomberg report, Epic has switched to an “AI-first” workflow on its unannounced Marvel–Star Wars-themed shooter with Disney: according to internal messages, there is no more hand-written code and the art is generated by a fully AI-driven pipeline; the release has been delayed indefinitely. CEO Tim Sweeney wrote on X that the goal is not to do the same with fewer people but to build more and better, and to bring the benefits to teams using Unreal too. An Epic spokesperson spoke of “AI-accelerated development” led by Epic’s developers. Source: Ars Technica (Bloomberg), Techmeme.
Anthropic: reasoning data extracted through distillation attacks fell by 90–99%. According to the company’s Monday report, new layers of protection — including “preserved thinking,” a two-stage classifier and limits on repeated attempts — have cut the number of reasoning tokens extracted by distillers on its own platform to roughly a tenth to a hundredth. Anthropic cited the development of Zhipu/Z.ai’s GLM-5.3 as an example, and said it has banned tens of thousands of accounts linked to distillation. The measures also involve trade-offs: some integrations had to be updated, and blocked requests may also be billed. Source: Anthropic Research.
What this means for you: The launch of Muse Wonder and the Maia 300 in-house chip strategy show that frontier capability and inference cost are decided together when you choose a vendor — if you are planning agents or your own models, our agent solutions and vendor-neutral consulting help make it clear what is worth committing to. Anthropic’s IoT replication and distillation reports, meanwhile, put security and oversight questions on the agenda; if you need a regulatory or governance framework, our governance consulting is at your disposal.