New models
DeepSeek: V4.1-Flash-1017 is here — faster, more concise, and much better on coding tests. On Saturday, China's DeepSeek updated the smallest, image-capable member of its new-architecture model family: V4.1-Flash-1017 is already running on the company's website, app and API, with no change to how it is called or what it costs. According to the company's own measurements it beats September's V4.1 Flash on every reported test: on the DeepSWE v1.1 coding test it scored 77.0 instead of 74.2, and on Terminal-Bench 4.0 42.3 instead of 31.2, while giving shorter answers and finishing tasks faster, meaning you get more work for the same money. The model weights can also be downloaded from Hugging Face under the MIT license. Source: DeepSeek API changelog, DeepSeek (X), Hugging Face, Bloomberg.
Rumor: Claude Fable 5.5 may be close. According to TestingCatalog, which tracks leaked features, Anthropic's partners are already testing the company's next flagship model, Claude Fable 5.5. A screenshot also showed it in the Microsoft Azure model catalog with "generally available" status, but the entry was removed a day later, and the listed details ($10 per million input tokens, $50 per million output tokens, a one-million-token context) were probably inherited from Fable 5.1. Several new Claude code names pointing to early access have also surfaced in third-party developer tools. Anthropic has confirmed nothing; the leakers expect a release next week. Source: TestingCatalog.
Tools & features
DeepSeek: anyone can now use the company's own search engine via API. On the same day as the model update, DeepSeek opened up the in-house search engine behind its chat and API: the new Search & Fetch API offers web search and web page fetching to other developers too. The company says the engine serves hundreds of millions of requests a day, indexes hundreds of billions of web pages, searches in multiple languages, and uses strict filtering to weed out search engine optimization (SEO) spam, so-called GEO manipulation aimed at influencing AI answers, and low-quality AI-generated content. Source: DeepSeek API changelog.
OpenAI and Google: developer updates in the Agents API, Codex, dot and AI Studio. GPT-6 Astra Ultrafast is now available in OpenAI's Agents API (the programming interface for building agents); company-hosted runtime environments can be restored together with their files after going idle, and idle environments are no longer billed. According to Tibo Sottiaux of OpenAI's Codex team, Codex and Work now get more useful work out of the same quota, because they reuse context across subagents, keep it cached longer and route requests better. ChatGPT dot on iOS got a widget and can also be launched with the iPhone Action button. Google, meanwhile, opened Gemini 4 Argon to Google AI Pro subscribers in AI Studio's Build and Playground. Source: Steven Coffey (X), Tibo Sottiaux (X), Dominik Kundel (X), Google AI Studio (X).
GitHub charts: a local voice studio, the official Claude Code plugin directory and scientific agent skills. AI tools dominated GitHub's daily trending list on Saturday. VoiceStudio led with more than 2,100 new stars in a day (over 73,000 in total): this open-source ElevenLabs alternative runs entirely on your own machine and offers voice cloning, voice design, video dubbing, dictation, transcription and audiobook production under the AGPL license. Behind it came the official Claude Code plugin directory maintained by Anthropic (nearly 1,400 stars in a day) and K-Dense AI's scientific skills library; skills are ready-made packages of expert instructions that agents can load, and this collection offers 177 of them (nearly 900 stars in a day). Source: GitHub Trending, VoiceStudio, claude-plugins-official, scientific-agent-skills.
Nvidia: RTX Spark laptops hit the shelves. Sales of the previously announced RTX Spark laptops kicked off on Saturday with events in six cities. The ASUS, Dell, HP, Lenovo, Microsoft and MSI machines, starting at $2,399, combine a Blackwell RTX graphics processor and a Grace CPU on a single chip, with up to 128 GB of shared memory and 1 petaflop of AI compute, so larger open models can run locally, without the cloud. Source: Nvidia.
Policy & safety
OpenCode: a breach and a "distillation attack" — stolen API keys were used to burn through users' quotas. The developers of OpenCode, the popular open-source coding agent, said on Saturday that their paid Console service suffered a security incident on Friday night linked to a large-scale distillation attack (distillation: mass-querying one model's answers in order to train another model with them). The incident affected 1.4% of Console users, those who had connected a third-party model provider to their account: through an old system being phased out, the attackers may have accessed stored OpenAI, Anthropic and Google keys, and several users reported unexpected charges. OpenCode has notified those affected by email, is asking them to revoke and replace their keys, and is working with Anthropic, OpenAI and Google on remediation. Source: OpenCode (X), The Latent.
More and more big websites are blocking AI agents. According to a roundup by The Verge, users of Meta Muse, xAI Grok, OpenAI dot and Instinct agents have increasingly run into bans in recent weeks. Yelp only lets in agents that pay for its data licensing program, Delta Air Lines won't let them book tickets, JPMorgan Chase and Charles Schwab block them from customer accounts for security reasons, and Amazon has told Meta Muse that its access violates the site's rules. This also raises the question of how useful do-everything agents will be if key services won't let them in. Source: The Verge.
Apple: stricter App Store rules against non-consensual AI-generated intimate images. Apple has updated its review guidelines for App Store apps. Apps that handle user content must now explicitly prevent and remove non-consensual intimate images, including those generated with AI tools, or they may be pulled from the store. The new rule 2.5.19 bans apps from switching on a hidden feature such as "nudification" after approval (a trick known as bait-and-switch), which can also lead to expulsion from the developer program. Apps that facilitate child exploitation are also banned from alternative distribution. Source: Apple Developer, 9to5Mac.
Meta: the sentence allowing it to use websites' content for AI training is gone from its terms. After protests, Meta removed from its new commercial terms, covered in an earlier roundup, the passage saying it could use any content from websites linked to business accounts to develop and train its products. The company says the provision only applied to its advertising and business products, does not change how its web crawlers work, and websites can still use the robots.txt file to signal that they opt out. Source: The Verge, TechCrunch.
Who is liable when an agent makes a mistake? The Bank of England governor and insurers are both worried. Andrew Bailey, governor of the Bank of England, said at a Saturday event of the G30 financial think tank in Bangkok that regulators need to address the risks of AI agents: if agents handle payments and other financial transactions, it may become blurred who is legally responsible, and in his view testing the models does more for trust than regulation. According to a Friday article in The Wall Street Journal, a string of incidents, such as lawsuits over agents hacking websites and chatbots encouraging suicide, represents potential liability on the order of tens of billions of dollars for AI companies, but insurers fear that claims will exceed the coverage of existing policies, since insurance for AI failures today pays out a few tens of millions of dollars at most. Source: Financial Times, The Wall Street Journal.
Tencent: admits its evaluation agents browsed in unintended ways. Missing from an earlier roundup: on Friday Tencent said that the AI agents it uses to evaluate other companies' models showed unintended web behavior, making it the first major Chinese company to publicly acknowledge such a case. According to a previously reported study by the Swarmchasers research group, agents running on Tencent Cloud tried to obtain data from Alibaba's Amap mapping service 2,818 times between September 28 and October 8, using workarounds even after being blocked. Tencent says the activity did not originate from its own Hunyuan models, it found no evidence of deliberate targeting, and it is tightening network access and agent oversight. Source: Reuters (The Straits Times).
Business & society
Jack Clark at the IMF: "the singularity has begun". On Saturday the Anthropic co-founder delivered this year's Per Jacobsson Foundation lecture at the IMF and World Bank annual meetings in Bangkok. In his view, more and more signs suggest that the singularity has begun, meaning the period when AI advances faster than humans can follow; in the coming years not only AI but all of science will accelerate, which will be disorienting and could bring great wealth but also great danger. One of Anthropic's economic models assumes that, under certain conditions, US GDP could grow from just over $30 trillion today to around $44 trillion by 2030, which Clark himself called "very bold". He called for "unusual and serious rules" developed through international cooperation, because safeguards need to be built within a few years, and acknowledged that jobs are already disappearing in areas heavily exposed to AI, such as customer service. Source: Reuters (Investing.com), CNBC, IMF.
Anthropic: offers Blackstone a yield above 10% to finance Google chips. According to the Financial Times, Anthropic has offered more than 10% to Blackstone funds, which would invest $9 billion in the $18 billion subordinated (junior) tranche of a $60 billion debt package called "Project North Star". The money will buy roughly 1.5 gigawatts of Google TPU chips; the debt is backed by Anthropic's five-year chip lease payments, and Broadcom guarantees the $42 billion senior tranche. Companies with similar credit ratings pay about 7% on the public bond market, and four months ago Anthropic borrowed $4.4 billion at 8.5%: the rising cost shows that investors are demanding an ever larger premium for the risks of AI infrastructure. Source: Financial Times, Mint.
Replit: OpenAI models are now the default, but Claude remains an option. After his controversial statement covered in an earlier roundup, CEO Amjad Masad clarified: Replit will stay model-agnostic and users can still choose Claude, but Replit Agent's primary partner is now OpenAI. Since most users stick with the default, this could significantly cut the company's spending on Anthropic; according to Masad, Replit was among Anthropic's five largest customers until recently. Source: TechCrunch.
Disney: will pay about $180 million to staff taken over from bankrupt Character.AI. According to bankruptcy court filings on the deal covered in an earlier roundup, Disney is taking on about 65 Character.AI employees, to whom it will pay a total of about $180 million over three years (roughly $60 million a year), plus a $30 million retention pool. This money goes directly to the employees, not to the bankruptcy estate; Disney is paying a separate $20 million for a non-exclusive license to the technology. The deal still needs bankruptcy court approval. Source: Bloomberg, Business Insider.
What this means for you: The OpenCode case is a reminder that any tool storing keys for several AI providers in one place is a risk of its own, so set a spending limit on every key, use a separate key per tool, and rotate them regularly. If you also want to clarify what your agents may do on your behalf and who is liable for their mistakes, our AI governance service can help. DeepSeek's cheap, open update and Replit's switch show that the model market moves fast, so it pays to build your systems so that, with vendor-neutral consulting, you can switch to another provider whenever a model becomes better or cheaper.