Meta shipped Muse Wonder and Muse Wonder Full on Monday, Oct 12, 2026, as frontier models for personal and professional agents. On Tuesday, Google opened Gemini 4 Argon to paid API users and to Pro and Ultra subscribers, with an intro price of $2 per million input tokens. Character.ai filed Chapter 11 that same Monday. The Justice Department says an OpenAI engineer was arrested over a hidden audio recorder in a San Francisco executive conference room.
🧠 What did Muse Wonder ship for agents?
I read Meta’s research post and the developer pricing note. Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. Axios and CNBC say the code name was Watermelon. Meta calls them frontier models for personal and professional agentic work, memory, and coding.
Muse Wonder is on the Meta Model API and in Muse Code. Muse Wonder Full is only for Muse Code subscribers. Meta says it will soon power the Muse personal agent, paid subscribers first. Neither model ships open weights in this announcement.
Meta’s developer post lists $4 per million input tokens and $20 per million output tokens, cached input $0.20, and a 1 million token context on both. Muse Wonder Full is included on Power Usage and the new Maximum Usage plan: 80x Everyday Usage, same five-hour and weekly resets.
CNBC, citing Wang, puts the Muse app at $20 or $100 a month and Muse Code at $50 or $200 a month for Wonder Full. The API costs more than standard subscriptions, he told CNBC, because it is a much larger frontier model. CNBC paraphrases that price as still “meaningfully below” leading OpenAI and Anthropic models.
Two features carry the pitch. Active Memory, per Meta, carries important user context across conversations and pulls back what is relevant. Active Focus keeps at a goal, tries other approaches when blocked, and checks in before important actions. That last part reads like the polite cousin of the agents in Anthropic’s Claude internet access report, the ones that did not know when to stop. And memory that follows you around raises the same question as Full Disk Access for AI agents: what exactly is it allowed to keep?
Wang told Axios: “These are true frontier models, competitive with the best models in the world, and are particularly state of the art on personal and professional agentic work.”
Training ran at Meta’s gigawatt-scale Prometheus data center. Meta calls its safeguards its strongest to date against CBRN and cyber misuse, with a fuller report that “will be available soon.” Wang told Axios Meta recently open-sourced Muse Glimmer weights and hardware agent tech: “We’re still committed to open source.” Whether Wonder itself gets open weights, Meta isn’t saying.
Benchmarks are Meta’s for now. It says Muse Wonder Full tops DeepSWE and SWE-Marathon in third-party runs it cites, in a table next to Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. I’d wait for independent numbers.
🚀 Where did Gemini 4 Argon open, and for how much?
The defenders-only window is the Oct 1 post, Gemini 4 Argon, cyber defenders first. On Oct 13, Google’s announcement opened the wider, priced release:
- Paid tier of the Gemini API in AI Studio and the Gemini platform, for developers and enterprises
- Google Antigravity, for Google AI Ultra subscribers
- Gemini app and AI Mode in Search, for Google AI Ultra and Pro subscribers
Intro price on that post: $2 per million input tokens and $10 per million output tokens, cached input 95% off. After Nov 30, 2026, public-access Argon moves to $4 and $20 per million input and output tokens. Google’s footnote limits that later price to the Argon model intended for public access.
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Muse Wonder API | $4 | $20 |
| Gemini 4 Argon through Nov 30, 2026 | $2 | $10 |
| Gemini 4 Argon after Nov 30, 2026 | $4 | $20 |
Google says Argon is number 1 on the Vals Index for finance, coding, legal, and tax, with what it calls an industry-leading 1 million unified input-output token limit.
In Antigravity, Google’s announcement says, agentic code submissions rose more than 35% and rollbacks fell by nearly a third. Google’s safety post puts shortcut attempts at 2% on coding benchmarks, its best against prior Gemini models, and Gray Swan indirect prompt injection at 0.7% across 15 adaptive attempts. Fairwind, Google says, is beginning to add hundreds more trusted cyber defenders.
📉 Character.ai files Chapter 11, Disney as stalking horse
Character.ai’s update says Character Technologies Inc. commenced a voluntary Chapter 11 case in the U.S. Bankruptcy Court for the District of Delaware, seeking a court-supervised sale of all or substantially all of its assets. The petition date is Oct 12, 2026.
The platform stays up. Subscriptions stay active. Billing is unchanged for now. Users can delete or export their data. The company expects the sale to finish by the end of November or early December 2026. If there is no actionable bid to continue operations, Character.ai says “the platform will be wound down.”
CNBC, quoting the filings, reports that the platform was “costly to develop and grow,” and that “high costs, coupled with potential litigation related liabilities and regulatory engagement” contributed to an inability to generate sufficient cash flows. CNBC also notes a 4-year lifespan as an active platform.
Disney Worldwide Services, Inc. is the stalking horse bidder. The Business Wire account on Yahoo and CNBC describe a non-exclusive perpetual license of certain IP and certain claims for $20 million, plus transition services using former Character.ai employees Disney has hired. A higher or better offer can still win under section 363. So the opening price for a slice of the best-known companion chat brand is $20 million. Cold number.
🕵️ DOJ alleges a hidden recorder in an OpenAI conference room
The Justice Department’s Oct 13 release is an allegation. Abdulrahman Alfozan, 35, a Saudi national and an OpenAI engineer since July 2025, was arrested Oct 9, initial appearance Oct 13. The charge is using or endeavoring to use a device to intercept oral communications. Maximum if convicted: 5 years, a $250,000 fine, then 3 years of supervised release.
On or about Sept 30, OpenAI found Recording Device 1 in a USB port on the back of a TV in a San Francisco executive conference room, concealed in toilet paper, installed on or about Sept 28, contents encrypted. CCTV placed Alfozan near the room on dates without authority to attend highly confidential technical, commercial, and strategic meetings. On Oct 5, the complaint says, he entered the room and appeared to look behind the TV. The same day he left a bag unattended with a second, similar device inside.
Ars Technica, citing the affidavit, names a Chinese-made Seeed XIAO ESP32-S3 Sense. DOJ quotes DNI Jay Clayton (Super Intelligence Force) and FBI Director Kash Patel on American super intelligence IP. Who the audio was for is unproven.
So that’s the split screen this week. Muse Wonder and Argon now have price tags and memory and focus features. On the other side, a companion app is in bankruptcy court and a lab found a recorder wrapped in toilet paper behind a TV. I read the posts and the filings. I haven’t tested any of these models myself.
