Meta Launches: A Personal AI Built Around Trust
Meta announced Muse, a personal AI agent that automates digital tasks—sending emails, booking travel, even selling your car—on your behalf. It's Meta's answer to viral competitors OpenClaw and Instinct, and the flagship product of Meta Superintelligence Labs, the AI unit Mark Zuckerberg formed last year to catch up with OpenAI and Anthropic.
Availability &
- **Live today:** Dedicated iOS/Android app, Muse.ai, and directly in WhatsApp; AI glasses support coming soon
- **Cost:** Free to try; an AI subscription plan required for heavy automation
Internally tested under the codename **"Hatch"** by Meta employees
What Muse Can Do
- Natural-language prompts, "no learning curve"
- Autonomously browses the web and operates third-party apps
- **Makes purchases** via Stripe's **Link**: single-use card numbers keep your real financial data off the internet; Muse is the first agent covered by Link's purchase protections, including **no-fee returns**
Architecture
- **Secure VM:** Every user gets an isolated virtual machine, separating user data from the agent's action-taking
- **Sentinel:** Monitors everything leaving the VM; matches actions to existing permissions or triggers human-in-the-loop approval
Approval prompts go **directly to users, not through the model**—a defense against prompt-injection attacks
- Already vetted by internal human and agentic red teams and a private bug bounty
Trust Problem
Muse needs deep access to personal data—something Meta's track record makes users wary of granting. Not:
- Meta is *policy-barred* from accessing user data, but **not technically blocked** (the standard Secure VM is not a locked box)
- Users can opt out of having their data used for
What's Next: Confidential VM In development with **Moxie Marlinspike** (creator of Signal), Confidential VM would add- A **trusted execution environment** with keys held locally by users—no one, including Meta, could access the VM
- **Third-party audits** by select security firms
- **Published binaries and a transparency log** so users can verify their connection
Bug Bounty Expansion
Muse now falls under Meta's public bounty program:
- Up to **$300,000** for valid vulnerabilities
- Up to **$130,000** for prompt-injection attacks a single user
Meta is late to the agent race but betting that aggressive security design— future user-controlled encryption—can win the trust its rivals haven't yet earned.
Another week, another batch of AI models. Last week, Anthropic, Google and Meta all released upgrades, while OpenAI rolled out GPT-6 Astra. The key question may not be who’s winning, but who can keep up. An AI insider dubbed the phenomenon "model fatigue" in a conversation with CNBC. The breakneck pace has not only fueled concerns about the dangers of increasingly powerful AI models, but also made technology decisions harder for businesses. Executives are devoting growing time and resources to assessing costs and capabilities as they seek to remain competitive.
September has had 4 business days. We have had 4 frontier model launches. Normal stuff.
Anthropic kicked things off Monday with Claude Fable 5.1. Then Google dropped Gemini 3.8 Flash. The funny thing about the name “Flash” is that it increasingly describes Google’s release cadence as much as the model. This is its 3rd Flash model in 6 weeks. 3.8 delivers better reasoning, coding and agentic performance while maintaining the speed and pricing of 3.7. Google seems increasingly uninterested in the leaderboard beauty pageant. Its game is cost and latency. That makes sense. It already has one of the largest distribution machines on earth. Gmail users do not particularly care whether Gemini can solve an obscure topology conjecture 3% more often than Claude. They care about speed, cost, and ubiquity. Less sexy, more lucrative.
Then Meta entered the group chat with Muse Spark 1.3, which explicitly optimizes for long-running agents that juggle multiple workflows, gather their own context, correct their plans, and keep going. It uses ~20% fewer tool calls and 25% fewer tokens than Spark 1.2 on coding work. The model is now firmly in the frontier cluster.
And then, because apparently nobody was allowed to get any work done this week, OpenAI dropped the big kahuna: GPT-6 Astra. They call Astra a “new generation of intelligence.” Greg Brockman went slightly further: “Welcome to the AGI era.”
Usually, statements involving the arrival of AGI deserve a fairly aggressive marketing discount. But the underlying numbers here are genuinely bananas: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, SOTA across computer use, browsing, software engineering, science, cybersecurity and professional work. Astra is great at operating browsers, navigating software, conducting research, manipulating complex environments, and completing multi-step tasks. The model is improving at the boundary between knowing what to do and being able to do it. It's a good model, sir.
Benchmarks aside, the Fable 5.1 versus Astra showdown will truly be decided over the next few weeks in the court of public opinion as thousands of users give both models ugly, ambiguous, real-world work. Then they will complain about both on X. Then they will slowly start using one more than the other. True PMF is measured in habits, not benchmarks.
Regardless of who comes out on top, the most encouraging part of this week is how many credible US labs shipped strong models within days of one another.
A year ago, it was easy to imagine the frontier collapsing into a two-horse race. This is a much more interesting market. The frontier is getting more crowded, the strategies more differentiated, and the crown more temporary. Exactly as it should be.
An underrated resource: when the AI labs release new models, they publish prompting guides that explain what changed and how to adjust your setup.
I read the guides for Astra and Claude Fable 5.1 this weekend (both linked in comments). One thing that surprised me was that several pieces of the guidance are exactly opposite:
- Fable 5.1 uses 𝘭𝘦𝘴𝘴 formatting than previous Anthropic models. Astra uses 𝘮𝘰𝘳𝘦 than older OpenAI models.
- Astra is 𝘮𝘰𝘳𝘦 likely to ask for your input. Fable 5.1 writes 𝘧𝘦𝘸𝘦𝘳 user-facing updates.
- Astra 𝘶𝘯𝘥𝘦𝘳-delegates to subagents. Fable 5.1 delegates 𝘢 𝘭𝘰𝘵, but the lead agent often just stops and waits for the subagents unless you tell it otherwise (classic manager behavior, tbh).
To be fair, each guide compares against that model's own predecessor (Astra vs. GPT-5.6, Fable 5.1 vs. Fable 5), not against the other.
But if you actively switch between frontier models, it feels really hard to tune your environment correctly to the differences in approach. A concrete example: a skill file that enforces heavy formatting might be exactly what Fable 5.1 needs, and cause over-formatting in Astra.
I suspect we'll start seeing teams keep different versions of their skill files for different models or harnesses. If anyone has tried this in their day-to-day setup, would love to hear how you managed it!