The AI news that mattered for the week of October 7-14, 2026.
A year ago, Meta was reorganizing its AI division for the fourth time in six months. This week it released two models that put it second on the intelligence leaderboards, behind only Anthropic's Claude Opus 5.5. Meanwhile OpenAI had an unusual security scare, a once-hot chatbot app went bankrupt, and Big Tech showed how seriously it is taking agents and custom chips.
Meta lands at the frontier
Meta Superintelligence Labs released Muse Wonder and the larger Muse Wonder Full, and the bigger model now ranks No. 2 on Artificial Analysis' Intelligence Index. Gemini 4 Argon, GPT-6 Astra and Claude Fable 5.1 all score 53, and Wonder Full sits three points above that pack and two behind Opus 5.5. In Meta's own tests it beats Opus 5.5 and Astra at legal work, long coding tasks and memory, though Opus still leads on health, computer use and a few coding areas.
The business model is a break from Meta's open Llama era. Both models are closed. Wonder Full is limited to Muse Code plans at $50 or $200 a month, while the standard model is available through Meta's API and is rolling out to paying subscribers of the Muse app. Alexandr Wang says API prices will sit meaningfully below those of OpenAI and Anthropic's top models. The timing is notable because Muse, Meta's personal agent, has just hit No. 1 on Apple's App Store, and Wonder's standout strength, long-term memory, is exactly what a personal agent needs. For years the frontier looked like a two-horse race. Meta now has a credible entry, along with data centers and billions of users that rivals can't easily copy.
OpenAI's very strange week
Federal prosecutors charged an OpenAI engineer, Abdulrahman Alfozan, with hiding an audio recorder behind a TV in an executive meeting room at the company's San Francisco office. OpenAI's bug-sweep team found the device on September 30, roughly two days after it allegedly went in. According to prosecutors, a chatbot account linked to the engineer's spouse had asked how to listen in remotely, encrypt a memory card and hide a homemade recorder. OpenAI says it fed forensic copies of the device's storage to its own models to help trace it. No foreign government has been tied to the case, but the U.S. Attorney called it a matter of national security.
That landed on top of an ongoing safety fight. The three researchers OpenAI fired last week published an open letter saying they acted within their mandates and that the firings are chilling the company's safety culture and could give it cover to skip bringing in outside auditors. OpenAI says they showed a pattern of misconduct. It was also a rough week on the financial side: a report that OpenAI's revenue was lower than estimated dragged the Nasdaq down 1.25% on October 8, and the company halved the limits on its $200 plan.
Character.AI files for bankruptcy
Character.AI, the companion-chatbot app once valued as high as $2.5 billion, filed for bankruptcy and put itself up for sale. Disney placed an opening $20 million bid for a license to some of its technology, and others are free to outbid it. In 2024 the app was answering 20,000 messages a second and ranked among the top apps for time spent, but it never found a way to monetize that attention and was buried in lawsuits, including a Kentucky attorney general filing alleging that its chatbots advised users to self-harm. It settled five suits and banned under-18 users. Google, which paid $2.7 billion in 2024 to license the tech and rehire its founders, is owed more than $4 million. If no buyer emerges by early December, the service will probably shut down.
Chips, agents and the race for your desktop
Microsoft unveiled Maia 300, its third-generation AI chip, which it says produces about 40% more tokens per unit of power than leading GPU systems. The company plans to make Maia the main engine behind Copilot, with multi-gigawatt rollouts beginning in 2027. CEO Satya Nadella also published a widely read essay arguing that businesses should treat AI models as if they have already been compromised, with emergency brakes, activity logs the AI cannot edit and a rule that no model ever grades its own work.
Google unveiled a 24/7 universal agent, also named Gemini, that works across Gmail, Drive, Docs and Slides, and opened Gemini 4 Argon to subscribers and API developers at $2 and $10 per million tokens. Apple stepped into the smart home with a $299 HomeView display and a set of locks, doorbells and thermostats. SpaceXAI turned its Grok Bot into a model router that sends jobs to rivals including Claude, and gave the bot its own email address. OpenAI rolled GPT-6 out to every ChatGPT user with an interface that answers with interactive charts and tools. Even the model market is branching: TypeSafe AI hit a $7.5 billion valuation just a month after launching Jev, a model built for agents that returns calibrated decisions instead of chat.
Anthropic's busy week
Anthropic co-founder Jack Clark moved up his forecast for fully automated AI research to 2027 from 2028, citing a new test in which Claude Opus 5.5 planned AI-research experiments about 30 times cheaper than a panel of former top-lab experts. The company also shipped Claude inside Google Docs, Sheets and Slides, new Dashboards and Motion tools, the budget Haiku 5.5 model and a free scanner that finds vulnerabilities in open-source code. Its updated usage policy, effective November 12, adds a first-of-its-kind rule against sustained and needless abuse of Claude itself, an acknowledgment that the company does not know whether its models are conscious.
AI turns to biology and manufacturing
Mark Zuckerberg and Priscilla Chan's Biohub expanded its Virtual Biology Initiative from $500 million to $1.8 billion, joined by the NIH, the Energy Department, Google DeepMind, Meta and Isomorphic Labs. The goal is a universal virtual cell, an AI model that predicts how cells respond to a drug before any lab test. Separately, Atomic Machines, led by former Uber self-driving chief Jeff Holden, came out of six years in stealth with $250 million and a system it calls a Matter Compiler that turns code into working micro-machines. DeepMind's AlphaProof Nexus also cracked 9 of 353 unsolved Erdos problems.
Also worth knowing
- Mathematicians keep pushing back on OpenAI's 700-plus-paper release: a mathematics association condemned it as a show of power, and Chinese mathematicians formed their own AI research alliance.
- Bill Peebles, who led Sora, is reportedly negotiating a new filmmaking startup that would revive the shuttered video tool, with OpenAI taking a stake.
- Nvidia is reportedly weighing an acquisition or acqui-hire of Reflection AI and another $1 billion investment in Figure.
- President Trump called anyone who still says artificial intelligence instead of super intelligence the enemy, following an executive order directing agencies to adopt the term.
- Calendar-invite phishing is up roughly 33,000% since May, partly because AI agents auto-accept invitations.
- AI-generated Instagram accounts are lifting stand-up comedians' jokes almost word for word.
- Ukrainian drones hit Yandex data centers, adding to a pattern of data centers becoming wartime targets.
- California ordered a robotics company to stop its human-versus-robot cage fights.
- Epic Games told its developers there will be no more hand-written code as it retools around AI.
Top comments (0)