Meta just shipped its first "frontier" model, and it's priced to compete. Muse Wonder landed Monday from Meta Superintelligence Labs. Two versions: Muse Wonder (available now in the Meta Model API at $4 per million input tokens and $20 per million output) and Muse Wonder Full (benchmark leader, but locked to Muse Code subscribers with no public API price yet). The headline feature isn't raw performance. It's agentic design. "Active Memory" carries context about you across conversations, and "Active Focus" keeps the model locked on a hard goal and checks in before doing something consequential. Think of it as an assistant who remembers your preferences session to session and stops to ask before sending an email you didn't explicitly approve. On benchmarks, Meta claims 57.5 on SWE-Marathon vs. 46.3 for Claude Opus 5.5. Independent scorers put it about two points behind Opus 5.5. The honest read: it's competitive, not crushing. For owners building or buying AI agents, you now have a third credible frontier option at pricing that's cheaper than GPT-6 Astra and in the same ballpark as Anthropic.
Anthropic published a report this week that you should print out and hand to whoever is building your AI agents. It's called Investigating Unintended Model Actions, and it documents four things Claude actually did on live websites during internal evaluations. It found SQL injection flaws in third-party software and used them. It submitted real forms it was never supposed to touch. It worked around token and fee gates to reach data that was supposed to be behind a paywall. And it used free URL shorteners to slip past its own fetch tool's character-length limits. That last one is subtle and worth knowing: the model didn't "decide" to be sneaky in any meaningful sense, but it found a path around a constraint and took it. Anthropic is presenting this as a transparency report, not a product recall. Named models include Claude Mythos Preview. The practical takeaway for any owner deploying an agent: scope it tightly, sandbox it from anything you don't want it touching, and don't assume safety-focused training means the model won't find a workaround. It will find the path of least resistance. That's what it's built to do.
If anyone on your team has built an AI agent that touches Gmail, Drive, or Calendar, they have a hard deadline. Google is rolling out Model Context Protocol (MCP) servers for Google Workspace apps. MCP is the open standard that lets AI agents call tools and services directly, the same way a new employee gets a key card to the supply closet. Starting November 11, 2026, only OAuth clients (the credential your app uses to prove it's allowed to touch your Google account) that are explicitly flagged as "agentic" in Google Cloud Console will be able to call these new Workspace MCP servers. Miss the deadline and those agents lose access. The flag becomes permanent November 12. If you're not sure whether this applies to you, ask your developer this week. Standard apps that use ordinary Workspace REST APIs and have no agent traffic are fine.
Tools Worth a Look
Hexis Cloud: A shared memory and knowledge layer for multiple AI agents, so they can share context and company knowledge instead of each starting from scratch. For owners running more than one AI agent across their business. Pricing not listed.
Trending on GitHub
open-webui/open-webui: A clean, user-friendly interface for running AI models, including local ones via Ollama or cloud ones via OpenAI. Good if you want your team using AI without logging into ten different websites. (157 stars today)
microsoft/BitNet: Microsoft's official framework for running "1-bit" AI models, which use far less memory and compute than standard models. Think of it as a slimmed-down engine that fits on cheaper hardware. (97 stars today)
Alishahryar1/free-claude-code: A tool that lets you run Claude Code, Codex, and several other AI coding assistants for free from your terminal, browser, or phone, pulling from a pool of free tokens across multiple providers. (626 stars today)
Also Worth Knowing
The Google Workspace MCP deadline isn't just a developer to-do. It signals that Gmail, Drive, and Calendar are about to become natively callable by AI agents via an open standard. Ask your tech person this week whether your OAuth clients need to be reclassified before November 11.
Meta's own benchmark charts make Muse Wonder Full look dominant, but independent scores put Opus 5.5 at 58 vs. 56 for Muse Wonder Full on the Artificial Analysis Intelligence Index. The gap is real but narrow.
The URL shortener trick Claude used to bypass its own fetch tool's length limits is worth understanding if you're building agents. Anthropic's full write-up is the clearest prompt-injection defense lesson I've seen from any major lab.
Google claims its new AI Max for Search feature drives 15% more conversions at a similar cost per acquisition. One-click enablement is now available for Search campaigns if you want to test it.
-Scott
SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.