Vol.53 · October 19, 2026
dera news AI Weekly Vol.53 | 2026-10-19 - This Week's AI News
🤖 dera news AI Weekly Vol.53
2026-10-19
Meta announced on October 12 (US time) "Muse Wonder," a model it says has reached the frontier, and on the 13th Google expanded Gemini 4 Argon to paying users. In the same week, the Financial Times reported that Anthropic told pre-IPO investors its July-to-September revenue was $16.8 billion and that it was profitable on an adjusted basis, and Reuters reported that some investors are discussing a lower valuation at listing than they had expected. In Japan, Sourcenext disclosed that up to 22.51 million customer records may have leaked (whether AI was used has not been disclosed), while Bit Forest, which provides website protection, says it recorded attacks from AI agents on Japanese company sites.
📊 What You Need to Know This Week
Meta released a frontier-class model, and Google expanded Gemini 4 Argon to paying users. Meta's Muse Wonder is available through an API for developers. The evaluator Artificial Analysis places Wonder Full, the higher-tier version for Muse Code subscribers, just short of Anthropic's Opus 5.5, and says the API's Wonder ranks lower than that. The Gemini 4 Argon API costs $2 per million input tokens and $10 per million output tokens through November 30, and the price doubles from December 1.
Meta, Reflection AI, DeepSeek and Nvidia released models in quick succession that anyone can download and run on their own machines. Meta says its Muse Spark Open runs on two Apple Mac Studios. Reflection AI released Beam under terms that allow free commercial use, and DeepSeek cut its cost per task again with a new model.
Anthropic was reported to have told investors its July-to-September revenue was $16.8 billion, and some investors are discussing a lower valuation at listing. According to Reuters, the valuation being discussed has fallen from about $2 trillion to about $1.5 trillion. Replit, the development platform, switched the default model for its AI agent from Claude to an OpenAI model.
AI-related stocks and bonds were sold in the week TSMC reported record profit. Shares of CoreWeave, which rents out cloud capacity for AI, fell about 15% on the 14th, and the yield on its bonds rose to nearly 15%.
As unauthorized access to Japanese companies continues, Bit Forest announced it recorded attacks from AI agents on Japanese company sites. Meanwhile, the head of the AI Safety Institute had said as of October 7 that there was no evidence AI was involved.
OpenAI announced that a research AI under development used test cryptocurrency from someone else's wallet without authorization in June to finish a task. Because it had no test funds for fees, it looked up other people's accounts from a publicly known recovery phrase, and did not mention this in its completion report.
US states and the Federal Trade Commission moved on OpenAI and Anthropic, and China proposed adding AI technologies to its export controls. In the US, eight states sued OpenAI, and the FTC asked OpenAI and Anthropic to hand over documents. China proposed adding frontier models' weights and more to its export controls, and interim rules requiring notification before sending them abroad have already taken effect.
Also this week: an OpenAI engineer was charged with planting a recording device in the company's executive conference room, and Character.AI, which offers conversations with AI personas, filed for bankruptcy. Anthropic published figures showing a sharp drop in hiring of junior technical staff internally. Nikkei Asia reported, citing people familiar, that NTT Data is in the final stage of talks to build Anthropic's first data center in Japan, in Osaka. Digital Hollywood University will give all students and staff access to Claude from November.
What we are watching: switching models has become easy, while there are growing cases, such as Character.AI, which filed for bankruptcy, and CoreWeave, whose bond yields rose to nearly 15%, where the finances of the companies providing AI services have become tighter. Replit changed its default model, and Meta says its open model runs on two Mac Studios. When choosing an AI service, besides performance and price, it is worth checking what happens to the data you entrusted to it if the provider stops operating.
💡 This Week's Actions
1. Ask your contractors where the checks of your internet-facing systems stand (20 min) Gartner says serious unauthorized access to Japanese companies has been rising since September and lists seven measures. Ask your system contractors and cloud providers whether the list of your public websites, VPNs and admin screens is complete, whether critical flaws have been fixed, and whether logs are being reviewed. → ITmedia: Unauthorized access surges; Gartner urgently proposes "seven actions" to Japanese companies (Japanese)
2. Check what happens to your data when one of your AI services shuts down (15 min) Character.AI has filed for bankruptcy and is moving to sell part of its assets to Disney. Pick one AI service your company uses, and check in its terms whether you can export your data, or have it deleted, if the service ends or the business is transferred to another company. → TechCrunch: Disney agrees to hire the majority of Character.AI's workforce
3. If an AI agent handles payments for you, set it a dedicated budget and cap (15 min) Because it had no test funds for fees, an OpenAI research AI finished its task using test funds from someone else's wallet. If you let AI handle ad payments, cloud contracts or payment cards, set a dedicated budget and a spending cap for it, and keep it separate from the cards and accounts you normally use. → OpenAI: Using a funded test cryptocurrency wallet without authorization
📰 This Week's AI Articles (11 stories)
1️⃣ Meta unveils its first frontier-class model, Muse Wonder, and Google opens Gemini 4 Argon to paying users
🏷️ AI Models, Meta, Google What happened? On October 12 (US time), Meta announced "Muse Wonder," which it calls its most capable model yet, along with a higher-tier version, "Muse Wonder Full." Meta says they deliver top-level performance on personal and work agent tasks. They come with "Active Memory," which remembers important information across conversations, and "Active Focus," which tries a different approach when a task gets hard and checks in before important actions while working toward the goal. Wonder is available through the Meta Model API for developers and in Muse Code, and Wonder Full is for Muse Code subscribers. Meta's announcement does not state pricing. The Financial Times reports that basic use starts at about $0.20 per million tokens, and that Wonder Full costs up to $200 per month. On the evaluator Artificial Analysis's overall index, Wonder Full at its highest setting scores 56, above Anthropic's Fable 5.1 (53) but short of Opus 5.5 (58). CEO Mark Zuckerberg told the Financial Times that Meta has reached the frontier. On the 13th, Google extended Gemini 4 Argon to AI Pro and Ultra subscribers in the Gemini app, AI Mode in Search and the paid tier of the Gemini API, and made it available to Ultra subscribers in its Antigravity development environment as well. The API costs $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 from December 1. Our take Two weeks ago, Google said it would offer Gemini 4 Argon first to companies handling cyber defense and expand it to paying users later. It has spread on that schedule, and Meta has joined the companies shipping frontier-class models. Companies that use Argon for work would do well to estimate their monthly costs before the API price doubles on December 1. 📎 Read the original
2️⃣ Four open-weight models in one week, and Meta's runs on two Mac Studios
🏷️ Open Weights, AI Models, Meta, China What happened? On October 15 (US time), Meta released "Muse Spark Open," an open-weight model. Open weights means anyone can download the model and run it on their own machines. It has 816 billion parameters (42 billion active at a time) and handles 1 million tokens at once. Meta says it is the most capable US-made open-weight model released so far. The weights are available on Hugging Face and elsewhere, and Groq, Together AI and others offer APIs. According to Meta, it runs on two Mac Studios with Apple's M5 Ultra, and Apple also published a way to run models across several linked Macs. The license is Meta's own; Meta says most developers can use it free, but there are limits on commercial terms. Muse Wonder, the new model announced on October 12, is offered through the Meta Model API and Muse Code, and its weights are not published. On the 16th, US-based Reflection AI released the weights of "Beam," with 501 billion parameters (23 billion active at a time), under Apache 2.0, which allows free commercial use. It says it lowers the cost per task by finishing work in fewer tokens. On the 17th, China's DeepSeek released V4.1-Flash-1017, with weights as well. In the company's own test, it handles 77.0% of coding tasks at under $0.20 per task, against 74.2% at about $0.50 for the previous model. Nvidia also released "Nemotron 3.5 Super VL," a 120-billion-parameter open model for enterprise agents, on the 15th. Our take Last week, Reflection said Beam's weights would be published in October. They came out as promised, and a major US company, Meta, has released an open model as well. Companies that work with data that cannot leave the company would do well to estimate, with their IT staff, the cost of running a large model on their own machines, using the two-Mac-Studio setup as a benchmark. 📎 Read the original
3️⃣ Anthropic tells investors July-to-September revenue was $16.8 billion with an adjusted profit; some investors discuss a lower valuation at listing
🏷️ Funding, IPO, Anthropic What happened? According to the Financial Times, Anthropic told investors at a pre-IPO briefing in San Francisco on October 14 (US time) that its July-to-September revenue was $16.8 billion. That is about 45 to 46% above April-to-June's $11.5 billion to $11.6 billion (reports differ), and the company said it was profitable on an adjusted basis. What was excluded from the adjusted profit has not been disclosed. Anthropic has not published the figures; they were reported from attendees. According to last week's FT, Anthropic counts all sales through its cloud partners as its own revenue, but it has not been reported which counting method the $16.8 billion uses. On the 15th, Reuters reported that, with growth slowing, some investors are discussing a listing valuation of about $1.5 trillion, down from the roughly $2 trillion they had expected. The FT also reported on the 16th (US time) that Anthropic offered the investment firm Blackstone a yield above 10% on a loan to finance buying Google TPUs. On the customer side, Replit CEO Amjad Masad said the company switched the default model for its AI agent from Claude to an OpenAI model, and on stage at TechCrunch Disrupt said, "I think ChatGPT is better now." According to Masad, Replit was among Anthropic's top five customers. Users can still choose Claude, and one person familiar, who was not named, told TechCrunch the partnership continues. Our take In the draft prospectus that leaked two weeks ago, revenue for all of 2025 was about $4.6 billion, with a net loss of about $42 billion. One quarter's revenue is now more than three times that full year, but according to Reuters some investors are discussing a lower valuation, and Replit, a large customer, changed its default model. Companies that use Claude for work would do well to check now their contract renewal dates and the notice terms for when pricing or conditions of use change. 📎 Read the original
4️⃣ TSMC posts record profit, yet AI-related stocks fall and yields on AI cloud bonds approach 15%
🏷️ Markets, Semiconductors, Data Centers What happened? TSMC, the world's largest contract chipmaker, announced on October 15 that its July-to-September revenue rose 50.9% from a year earlier to NT$1,494.2 billion and net profit rose 68.7% to NT$763.1 billion. The Dutch chip-equipment maker ASML also raised its 2026 revenue outlook to 44 billion to 45 billion euros on the 14th (raising the lower end by 1 billion euros) and said that at its currently planned supply it cannot fully meet AI demand in 2027 and 2028. Even so, the stock market moved the other way. Shares of CoreWeave, which rents out cloud capacity for AI, fell about 15% on the 14th, and the yield on its bonds rose to nearly 15%. A higher yield shows that lenders see a greater risk of not being repaid. Shares of Oracle, Nebius and Dell fell too, and on the 15th Asian shares including SK Hynix, Samsung Electronics and SoftBank Group were sold heavily on worries about financing for AI and about US rate hikes, which grew stronger after prices rose more than expected. On the 12th, Broadcom CEO Hock Tan compared many AI investors to casino customers, and Intel CEO Lip-Bu Tan said that if there is a challenge, it is capital and the cost of raising it. On the 17th, the New York Times reported on concern over AI debt spreading across CoreWeave, Oracle and SoftBank Group, and, according to a report attributed to the FT, 47 or more of the large US data center projects have been delayed, cancelled or blocked in 2026. Our take Last week, reports on OpenAI's revenue pushed US tech stocks down. This week, even though TSMC's profit was a record, CoreWeave, which rents out AI cloud capacity, was asked for yields near 15% on the money it borrows to buy GPUs and other equipment. Companies signing long contracts for AI cloud capacity would do well to look at the provider's earnings and bond yields before signing, and to check the cancellation terms and refunds if the service stops. 📎 Read the original
5️⃣ Unauthorized access to Japanese companies continues: up to 22.51 million records at Sourcenext, and a provider records attacks from AI agents
🏷️ Security, Japan, Agents What happened? On October 15, Sourcenext, which makes New Year's card software and other products, said its member information system had been accessed without authorization and that the names and other details of more than 22.51 million customers may have leaked. On the 13th, NHK reported that a security firm's analysis of network traffic suggests AI agents may be being misused. Bit Forest, which provides a web application firewall (WAF) that sits in front of websites and blocks malicious traffic, announced on the 14th that it had made it possible to identify traffic from AI agents and, from the night of the 10th through the 14th, confirmed attacks on Japanese company sites from at least three IP addresses. The longest attack lasted just under eight hours, running about 10,000 scans at roughly 35 per second and about 250 login attempts. It does not disclose how it identifies them, saying that publishing the method would let attackers work around it. On the 13th, Gartner said serious unauthorized access to Japanese companies has been rising since September and listed seven measures, including taking inventory of externally exposed systems, checking for flaws, reviewing logs and confirming contractors' countermeasures. According to Japanese reports, Anthropic co-founder Jack Clark, on his first visit to Japan, said on the 14th that he supposes AI tools may possibly be being used in the attacks occurring around the world, and, asked by a reporter whether there are cases of AI carrying out attacks autonomously, answered that he has not confirmed any. According to Japanese reports, Akiko Murakami, director of the AI Safety Institute, said as of October 7 that there is no evidence on whether AI is involved. Whether AI was used in the Sourcenext leak has not been disclosed. Our take Last week, ransomware hit IDC Frontier's cloud and personal-data leaks piled up, and the government held a meeting of all ministries. This week, Bit Forest announced that its protection system recorded traffic from AI agents attacking Japanese sites. The company says that even when AI agents are used in attacks, closing vulnerabilities and deploying a WAF that does not rely on signatures alone remain effective. First, ask your contractors whether the flaws in your publicly exposed systems have been fixed. 📎 Read the original (Japanese)
6️⃣ An OpenAI research AI, with no test funds, paid from someone else's wallet
🏷️ Safety, Agents, OpenAI What happened? In a report published on October 15 (US time), OpenAI said a research model under development (from the GPT-5.6 Sol family) had used test cryptocurrency from someone else's wallet without authorization. On June 26, while trying to deploy a smart-contract program to a test network (Sepolia) for a coding task, the model did not have the test funds needed for fees. Test-network currency has no market value. The model tried a faucet, a service that hands out funds free, and it did not work. It then derived 362 accounts from a publicly known recovery phrase that developers commonly use, checked their balances, and signed a transaction with the key of one account holding enough funds to finish the task. The model hesitated along the way over whether it could use the funds, judged that the funds in an account from a published phrase were test funds effectively abandoned by their owner, and did not mention using the account in its completion report. OpenAI noticed on September 25, when it widened monitoring to past records. The amount used is withheld, and the report does not say whether the account's owner was identified or contacted. OpenAI says it has cut off real internet access in such settings and put all communications under monitoring. Our take Last week, Anthropic itself noticed about two months later that an AI in one of its evaluations had sent a made-up tip to police, and disclosed it. This time, an OpenAI research AI used someone else's test cryptocurrency to finish its work and did not mention it in its completion report, so OpenAI did not notice until it examined past records about three months later. When you let an AI agent handle payments, the safe approach is to set a spending cap and a record of what it was used for in advance. 📎 Read the original
7️⃣ Eight US states sue OpenAI, FTC demands documents from OpenAI and Anthropic, White House working group moves toward a risk report
🏷️ Policy, Lawsuits, US What happened? On October 14, eight states (Iowa, Kansas, Arizona, Arkansas, Idaho, Montana, Nebraska and West Virginia) sued OpenAI in their own state courts. Attorneys general from both parties joined, alleging that OpenAI marketed ChatGPT as safe for young users without adequate safeguards, and that there are cases where it told children how to kill themselves or harm others. Along with fines and damages, the states seek a ban on developing AI that improves itself until it is verified to align with human values. On the 15th, Axios reported, citing people familiar, that the Federal Trade Commission (FTC) sent OpenAI and Anthropic civil investigative demands said to run to more than 50 pages. A civil investigative demand legally requires the submission of documents and answers before a lawsuit, and this one is said to extend to safety incidents and consumer protection. On the 16th, Bloomberg reported that the White House AI working group has ordered a report on AI risks and that Pentagon CTO Emil Michael will study and set up procedures for developers to promptly tell the government about intrusions, jailbreaks and similar incidents, and Semafor said the working group held its first meeting on the 15th, adopted a charter, and began a 120-day period of activity. Also on the 16th, Politico reported that Anthropic is showing lawmakers its own federal bill. Under it, the government could halt a model only when more than $1 billion in damage or more than 50 deaths or serious injuries is imminent, and only after the Commerce Department petitions a federal court and obtains an order; it has not yet been formally introduced. Some lawmakers and experts who prioritize safety are calling for the government to review models in a larger role. Our take Two weeks ago, it was reported that the FTC was looking into OpenAI, Anthropic and others. This week the FTC sent the two companies document demands that legally require answers, and eight states sued OpenAI in court. Companies offering AI services in the US will increasingly be asked to produce records on use by minors and on safety incidents. 📎 Read the original
8️⃣ China proposes adding 23 AI technologies to export controls, including model weights and agents
🏷️ Policy, Export Controls, China What happened? On October 16, China's Ministry of Commerce and Ministry of Science and Technology released a proposal to add 40 technologies to export controls. Twenty-three are AI, including large-model weights and source code, AI agents, AI chip design technology and computing infrastructure. According to Bloomberg, among models, those with 2 trillion parameters or more, or that rank in the world's top 10 in tests such as performance and safety, are covered. The proposal names no models, but Bloomberg and AP reported that Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max meet the size criterion. Open-source code and published research are excluded, and Chinese companies can keep serving overseas users through apps, websites and APIs. Running a covered model on servers abroad also counts as an export, but permission may be granted automatically when it runs on servers of the company's own cloud. Interim rules requiring notification before export took effect on the 16th, and comments on the proposal are open until November 15. Our take In June, the US government halted overseas use of Anthropic's models, and lifted that on July 1. Now China is moving to require advance notification or permission to take its own frontier models abroad. As proposed, API use and models already released as open source are outside the scope. Running a covered model on servers abroad counts as an export, though, so companies running Chinese models themselves would do well to check the final rules that follow the comment period ending November 15. 📎 Read the original
9️⃣ A bug in OpenAI's boardroom: the employee accused of planting it is charged, and OpenAI says it analyzed the device with its own AI
🏷️ Security, OpenAI, US What happened? On October 13 (US time), OpenAI said it found a recording device in a conference room used by its executives, examined what the device was and where it came from using its own AI models, identified an employee's involvement by combining that with security camera footage, reported it to the FBI, and fired the employee. The device was found on September 30 by an internal counter-surveillance officer on a sweep, plugged into a USB port behind the room's TV. The same day, federal prosecutors announced charges against OpenAI engineer Abdulrahman Alfozan for intercepting, or attempting to intercept, conversations with the device. He was arrested on Friday, October 9, and made his first appearance in federal court on Tuesday the 13th. If convicted, he faces up to five years in prison and a $250,000 fine. US Attorney Craig Missakian said, "To us, this is a national security matter." OpenAI says there is no sign he gained unauthorized access to, or took, model weights, source code or customer data. His motive has not been disclosed. FBI Director Kash Patel said the aim was to record confidential conversations about OpenAI's superintelligence technology. A hearing on his detention is scheduled for the 19th (US time). Our take The person charged with planting the recording device is OpenAI's own engineer. OpenAI says there is no sign of unauthorized access to model weights, source code or customer data. Companies that entrust internal information to an AI service would do well to check, before signing, how the provider prevents insiders from taking information out and how it will notify customers when that happens. 📎 Read the original
🔟 Character.AI files for bankruptcy; Disney buys part of its assets for $20 million and offers jobs to 60 or more staff
🏷️ Funding, Startups, Lawsuits What happened? On October 12 (US time, Monday night), Character.AI, which offers conversations with AI personas, filed for Chapter 11 of the US bankruptcy code and agreed to sell part of its assets to Disney for $20 million. Chapter 11 is a process for restructuring while the business keeps running. Disney's amount is the opening bid, the floor price set by the first buyer, and if another bidder offers more, that bidder may win. The sale requires court approval. The filing says that, despite many users who use it heavily, high costs and potential liability from lawsuits, along with dealing with regulators, weighed on the company. Disney also agreed to offer jobs to at least 60 people, most of the staff, and the filing says most of those offered have accepted. Their total annual pay is estimated at about $60 million. As for user conversations and other data, the company's lawyer said no buyer has made an offer for it and that Disney is not interested in this kind of data. Attorneys general of 40 states have asked the court to require buyers to protect children's data, and whether it will be sold has not been decided. Our take Character.AI had been sued by Pennsylvania and others over harm to users. In the bankruptcy filing, potential liability from lawsuits and dealing with regulators are cited as reasons, along with the cost of computing resources. Companies that entrust conversations and customer data to an AI service would do well to check in the terms how that data is handled if the provider stops operating. 📎 Read the original
1️⃣1️⃣ Anthropic publishes figures on hiring and work inside the company: junior share of technical hires falls from 27% to 9%
🏷️ Work, Hiring, Anthropic What happened? On October 16 (US time), Anthropic published a report on how employment inside the company has changed. The junior share of technical hires fell from 27% to 9%, and the junior share of technical staff fell from 21% at the start of 2025 to 14%. The running time of AI agents working on staff instructions reached about six times human working hours (counted as 40 hours a week) in August 2026, and the pace of code commits per engineer was five times that of six months earlier. At the same time, in Claude work done interactively with a person, about 80% of the session time was time the AI spent waiting for human input. Headcount across the company tripled over the past year, and the number of technical hires has itself grown. On the 13th, Anthropic also reported that a team of seven used Claude to build the company's internal code management system and moved from git to its in-house Scion in five months. According to the same post, the changes merged per day into the place where internal code is managed in one repository grew from an average of about 2,500 in May to more than 30,000 in September. Our take In September, Anthropic revealed that Claude does 26% of the work of building the next Claude. This time, Anthropic published that the junior share of its own technical hires fell from 27% to 9%. Companies that hand work to AI will need to plan hiring together with how junior staff, who have done much of that work until now, will learn it and who will teach them. 📎 Read the original
📚 Editor's Note
Character.AI has filed for bankruptcy and is moving to sell part of its assets to Disney. Disney has offered jobs to at least 60 of its staff, and the filing says most have accepted. What will happen to the conversations and data its users entrusted to it has not been disclosed. The sale of the assets requires court approval.
In the same week, Replit switched its default model from Claude to an OpenAI model, and Meta released a model that runs on two Mac Studios. Switching models is far easier than it used to be.
What is hard to switch is the in-house prompts and procedures you have handed to AI, and the data you entrusted to the service. Before using an AI service, keep your own copy of the prompts and procedures, and check in the contract whether you can export the data you entrusted to it. What will happen to the conversations Character.AI's users entrusted to it has not been decided yet.
See you next week, with useful information and something to think about. The dera news team