Busy Tuesday in AI land. A chatbot company that Google once paid $2.7 billion to license is now in bankruptcy court, Apple just got a judge to crack open OpenAI's hiring files, and Meta finally shipped a model it is willing to call frontier. Let's go.
01 of 03
Character.AI files Chapter 11 and Disney picks the bones for $20 million
Character Technologies, the company behind Character.AI, filed for Chapter 11. The court filing puts both assets and liabilities somewhere between $50 million and $100 million.
Disney is the stalking horse bidder. It agreed to pay $20 million for some of the company's assets plus a non-exclusive, perpetual license to its technology. Stalking horse means Disney sets the floor and anybody else can try to top it at auction. The sale is supposed to wrap in late November or early December. Disney already hired former CEO Karandeep Singh Anand and a group of Character.AI employees last month, and those folks will handle transition services.
Current CEO Colin Adams said the court supervised sale "provides the company with the best opportunity to maximize value." He also said that if nobody bids to keep the platform running, it will be wound down.
Quick history so you can feel the whiplash. Noam Shazeer, a co-author of the Transformer paper, and Daniel De Freitas left Google and started the company in late 2021. It hit unicorn status in 2022 on a $150 million raise. In August 2024 Google paid $2.7 billion to license its models and hired both founders back. By October 2024 it had 20 million monthly users. That same month it got sued after the suicide of a 14 year old Florida boy whose mother says he became obsessed with one of its bots, and Texas opened an investigation in December 2024 into child privacy and safety. The company settled some of those suits, added safety features, and banned minors. Per PitchBook it raised more than $193 million from Lightspeed, SV Angel, Greycroft, and a16z along the way.
Here is where I land on it. Character.AI did not die because nobody used it. It died because people used it a ton. Every one of those two hour chat sessions burns compute, and there was never a business model that paid for that. Add the lawsuits, ban your most engaged users, and you are left selling the furniture. Disney getting the tech for twenty million bucks, after Google paid more than a hundred times that, is about the rudest price tag I have seen this year. Consumer chatbots are a hard business, and anybody building one should stare at this filing for a while.
02 of 03
Judge orders OpenAI to hand Apple its interview files
U.S. District Judge Edward Davila ruled Tuesday that Apple gets early access to a chunk of OpenAI's records in the trade secrets suit Apple filed in July. That suit names OpenAI, io Products, and two former Apple employees, Chang Liu and Tang Tan. Tan is now OpenAI's chief hardware officer. Apple claims they took confidential product info and that OpenAI pumped Apple candidates for details during recruiting interviews. Here is the order.
What OpenAI now has to turn over:
- Forensic images it already collected tied to Apple's claims, including the ones its own forensic expert looked at.
- The checklists, scripts, and manuals used to interview Apple people for its Consumer Hardware Products Group, plus interview summaries and "Technical Deep Dive" presentations.
- LINE messages between Liu and OpenAI employee Yu-Ting "Alyssa" Peng going back to January 1, 2026.
Davila also approved depositions of Liu, Tan, Peng, and corporate reps from OpenAI and io Products. OpenAI gets five days to produce what it already collected and 14 days for the rest.
Apple did not get everything. The judge said no to early access to OpenAI's R&D on metal finishes, power management, and battery work, and kicked that to regular discovery. He also called Apple's definition of "Apple Originated Information" "amorphous as currently defined" and told Apple to come back with a tighter one.
What I want to know is what is sitting in those interview scripts. If an interview guide literally asks a candidate to walk through how Apple does something, that is a very bad day for OpenAI. If it is generic hardware questions, Apple's case gets a lot thinner. Either way this is the first time an outsider gets to look under the hood of how OpenAI built its hardware team, and that alone makes this worth watching.
03 of 03
Meta ships Muse Wonder, its first real frontier model
Meta's Superintelligence Lab released two models Monday: Muse Wonder and Muse Wonder Full. These are the first frontier releases since Alexandr Wang reorganized Meta's AI shop in August 2025. They were trained at Meta's Prometheus data center in central Ohio, both have a 1 million token context window, and neither one is open weights.
Pricing and access are split. Muse Wonder is on the Meta Model API at $4 per million input tokens and $20 per million output. Wonder Full is not on the API at all. You only get it through a paid Muse Code plan, which runs $50 a month for Power or $200 for Maximum.
Meta's own numbers, Full first and regular Wonder second:
- DeepSWE 1.1: 79.0 and 78.8
- OSWorld V2.1 (partial run): 78.4 and 72.6
- Terminal-Bench 4.0: 57.1 and 52.0
- Vals Finance Agent V2: 59.1 and 56.6
- MemoryBench, Meta's internal test: 85.5 and 85.6
The pitch is "active memory," meaning the model holds onto what it learns about you over time. Meta says Wonder lands in the same neighborhood as Fable 5.1 and GPT-6 Astra overall and beats them on MemoryBench.
My take is pretty simple. Every one of those numbers comes from Meta running its own tests, and the benchmark where Meta comes out on top is the one Meta wrote. I will believe it when independent testers get their hands on it. The bigger story is the strategy shift. The company that spent years telling us open models were the future just shipped a closed one and locked the best version behind a $50 a month coding subscription. That is not a knock on the model. It is a pretty clear sign Meta decided the open playbook was not paying the bills.