New models
pplx-embed-v2-context: Perplexity releases open-weight embedding models that handle both images and text. Perplexity's developer team published the new embedding models in two sizes, with 9 billion and 0.6 billion parameters; they turn documents into searchable numeric vectors. They return either a single compressed (int8) vector per document or one per chunk, with every chunk “seeing” the full document, which can make search systems more accurate. The paper on the models was co-written with the database company turbopuffer. Source: Perplexity Devs (X).
Specialized models: Hume Decision API draws conclusions from voice recordings, Nace AI's Drex model screens suspicious card payments. According to TestingCatalog, Hume AI launched the Hume Decision API, a voice-based model that analyzes the recording itself to answer the user's own questions, such as whether a caller sounds frustrated, and gives an estimate for the answer. Nace AI introduced Drex for Finance, which takes a document or transaction log plus a fixed list of answers and returns a probability in a fraction of a second, for example on whether a card payment is fraudulent. Source: TestingCatalog: Hume (X), TestingCatalog: Nace AI (X).
Tools & features
Claude API: beta message threads, so long conversations only need to send the new messages. On Monday Anthropic made message threads in the Messages API available in beta: the API stores the conversation on the server, so each follow-up request only sends the new messages and the ID of the previous message, and Claude starts responding sooner. This matters most for long conversations and agent loops. A thread expires 5 minutes after the last request and is deleted no later than one hour after last use; besides the Claude API, the feature is also available on AWS, Google Cloud and Microsoft Foundry. Source: Claude Platform release notes, Claude Devs (X).
DeepSeek Harness: Creator Mode for building your own reusable plugins. DeepSeek released Creator Mode for Harness, its interface for agentic work, which lets users quickly build their own plugins (extensions) for various tasks, such as a 3D model viewer or a video editing workspace, according to Bloomberg's examples. The feature strengthens the Chinese company's agent offering. Source: Bloomberg, TestingCatalog (X).
ChatGPT: the merger of the Chat and Work tabs begins in the desktop app. Following the announcement covered in our previous roundup, OpenAI told users in an email that it is merging the Chat and Work tabs in desktop ChatGPT, starting with selected users to gather feedback. In the unified interface you can bring in files and plugins, and ChatGPT can build websites and presentations or complete tasks on the web. Codex stays a separate surface but draws from a token allowance shared with ChatGPT, so some users will have to wait for their allowance to refill more often. Source: TechCrunch.
Dots: GPT-6.1 Terra joins the models powering OpenAI's always-on agents. According to Thibault Sottiaux, who leads Codex at OpenAI, the range of frontier models running Dots agents is expanding: GPT-6.1 Terra is being added alongside GPT-6 Astra, with more models to follow. Terra is faster and has a friendlier style, and it could help Dots reach more users. Source: Thibault Sottiaux (X).
Developer tools: a Runway plugin for Cursor, Qodo code review in Codex, and OpenAI's own Git backend. Runway's plugin gives Cursor everything it needs to build Runway's video API into an app with a single install. Qodo brings a company's internal knowledge and coding rules into Codex, so code review catches the most serious bugs before a pull request is even opened. OpenAI also unveiled Pando, its Git-compatible backend built for the development pace that Codex enables; the post also reached the front page of Hacker News. Source: Runway (X), OpenAI: Qodo, OpenAI: Scaling Git.
Hyra 2.0: Tencent's research agent now builds tools and workflows as well as solving tasks. Tencent's Hunyuan team unveiled version 2.0 of its research agent, Hunyuan Research Agent. According to the company, Hyra 2.0 no longer just solves problems but also builds reusable tools, models and workflows that speed up later research. Source: Tencent Hunyuan (X).
Policy & safety
Anthropic: according to its amended IPO filing, the company may soon reach the threshold of self-accelerating AI progress. According to the amended IPO prospectus (S-1/A) Anthropic filed on Monday, internal estimates and expectations about future models mean the company may “soon” reach the threshold of “dramatically accelerated, AI-attributable progress” set out in its Responsible Scaling Policy. This is so-called recursive self-improvement (RSI), where AI increasingly develops the next models itself. The company adds that none of its current models has crossed the threshold, that such an acceleration might only be recognized in hindsight, and that for the foreseeable future it does not consider it safe for Claude to design the next generations without meaningful human oversight. Source: TechCrunch.
New York Times: sues the Brave browser, saying it sells stolen content to AI companies. On Monday the New York Times sued Brave in federal court in Manhattan, alleging that the company collected the paper's copyrighted articles without permission and passed them on to Anthropic, Perplexity, Mistral, xAI and Cohere. According to the complaint, Brave built an index of more than 30 billion pages with an unblockable crawler, its browser bypasses paywalls, and it blocks publishers' ads while selling its own. The Daily News, the Chicago Tribune and the Orlando Sentinel filed similar suits. Source: TechCrunch, The Verge.
Claude Code: Anthropic confirms the model no longer copies specific, existing video games. Anthropic added two sentences to the system prompt of Claude Code running on Opus 5.5 so that the model does not rebuild other people's specific games, because infringing intellectual property violates its usage policy. Original games, games that fit a genre, and games built from the user's own designs and artwork can still be made. The decision responds to a wave of AI-made remakes and ports of old games. Source: The Verge, Tom's Hardware.
Wired: around 400 harmless-looking AI apps hide deepfake features that enable sexual abuse. According to a Wired investigation, around 400 apps and websites linked to Chinese-speaking developers present themselves as everyday AI tools, such as pet portrait makers, but hidden features let users create non-consensual sexual and violent deepfake images and videos. The network also gets around app store review. Source: Wired.
OpenAI: bigger bug bounties, another internal security sprint, and a national security adviser. OpenAI raised the rewards in its security bug bounty program and introduced extra bonuses for high-quality reports, for returning researchers and in priority campaigns. It also announced that, together with Cathedral Systems, it is preparing an AI-assisted internal security test series (security sprint) to harden its own systems. Dan Driscoll joins as a senior adviser working on national security modernization, cyber resilience and cooperation with US allies. Source: OpenAI: bug bounties, OpenAI: security sprint, OpenAI: Dan Driscoll.
Chinese open models: Western startups edit the weights to remove censorship and security gaps. According to Rest of World, Hirundo, Lazarus and other companies use machine unlearning and targeted weight editing to adapt Chinese open models to the expectations of their Western enterprise customers: they strip out censorship and state-aligned bias and make the models more resistant to cyberattacks. Source: Rest of World.
Business & society
Anthropic: quarterly revenue up twelvefold, but with a $23.3 billion loss. According to the amended IPO filing, Anthropic's third-quarter revenue was nearly $17 billion, twelve times as much as a year earlier, but its net loss grew to $23.3 billion from $12.1 billion last year, and operating expenses rose 292% to $17.5 billion. The company disclosed $895.4 billion in purchase and lease commitments, almost entirely for compute and infrastructure, and it could raise up to $3.5 billion from Japanese investors in the IPO. According to the Wall Street Journal, institutional investors see a valuation of $1.5–1.6 trillion as realistic, rather than the $2 trillion mentioned earlier. Source: CNBC, Bloomberg, The Decoder.
Marvell: plans to buy Lumentum in a $115 billion stock deal to grow in optics for AI data centers. Marvell would acquire optical component maker Lumentum in an all-stock deal that values Lumentum at an enterprise value of $115 billion. It is the largest acquisition of the year, and the aim is to combine electrical and optical technologies to take on Broadcom in the AI data center interconnect market. On the day of the announcement Marvell's shares fell 13% and Lumentum's nearly 19%. Source: CNBC, Reuters.
Grok Bot: tighter usage limits, as Musk says serving costs were “unsustainably high”. Over the weekend several Grok Bot subscribers found that their weekly allowance had unexpectedly run out completely. In a Sunday reply on X, Elon Musk confirmed this was not a bug, because the cost of the service was unsustainably high relative to revenue. SpaceX has not yet said how the limits will work in the future, even though only a few weeks ago Musk called Grok Bot the fastest-growing product he had ever seen. Source: TechCrunch, The Decoder.
Research: GPT-6 Astra improved a small Qwen model at math by directly rewriting 0.016% of its weights. Researchers say large language models can directly read and write the weights of other neural networks. In the experiment, without conventional training, GPT-6 Astra identified a flaw in the weights of Qwen3.5-0.8B that caused repetitive reasoning loops, and by rewriting 0.016% of the weights it raised the score on the MATH-500 math test from 7.9% to 48.1% and tripled it on the instruction-following test. The authors call this weight-space programming; the post also reached the front page of Hacker News. Source: Weight-Space Programming.
New benchmarks: the top model scores only 12% on SemiAnalysis GDPx, while OpenAI measures everyday tasks. SemiAnalysis's GDPx test measures AI agents on 206 hard tasks across 11 economically important fields, such as chip design, software development, finance and biology; even the leading Claude Opus 5.5 scored only 12%. OpenAI's new LifeVal benchmark, based on feedback from thousands of ChatGPT users, measures how much AI helps with everyday chores, from finding childcare to finances to government paperwork. Source: SemiAnalysis: GDPx, OpenAI: LifeVal.
Goldman Sachs: traders and bankers can no longer use advanced AI coding tools. According to Bloomberg, under Goldman Sachs's internal guidelines from October most employees lost access to tools such as Claude Code and GitHub's AI tools, and only engineers can keep using them. A bank spokesperson cited productivity and risk management and said the pilot phase had ended. Source: The Decoder.
Layoffs: citing AI, SUSE cuts 10% of its staff and Squarespace 19%. In a letter to its roughly 1,900 employees, SUSE's CEO wrote that automation and AI are fundamentally changing how software companies work, so the company is cutting headcount by about 10%. Website builder Squarespace is letting go of around 300 people, 19% of its staff; according to TechCrunch, the industry is being reshaped by AI tools that build pages from text prompts. Source: The Register, TechCrunch.
Investments: $45 million for Trajectory, $62.5 million for 42AI, $75 million for Tutor Intelligence. Trajectory, which offers models that keep learning from expert feedback without any programming knowledge, raised a $45 million Series A at a $300 million valuation, led by Sequoia with participation from Nvidia. Tel Aviv-based 42AI came out of stealth with $62.5 million in seed funding and a “behavioral world model” that predicts human decisions as probabilities. Tutor Intelligence, which builds warehouse robots, closed a $75 million Series B led by a16z. Source: TechCrunch: Trajectory, TechCrunch: 42AI, Forbes: Tutor.
OpenAI: Grid Protectors program so future data centers draw less power when the grid is under strain. OpenAI's US program, run together with utilities, aims to have future compute campuses reduce their power draw when the grid is overloaded, supporting the security of supply and keeping residential electricity prices from rising. Source: OpenAI, The Verge.
Oracle: raises $20 billion in stock and bonds for its data centers. Oracle is issuing $10 billion in common stock and $10 billion in bonds to fund its data center buildout and chip purchases; its shares fell more than 5% after the announcement. Source: Wall Street Journal.
Apple: plans an AI-powered home robot for as much as $1,000 by 2027. According to The Information, Apple's home robot, code-named J595, would move around the home on its own, handle FaceTime calls and switch on the lights when someone enters a room. The company is targeting 2027 for the device priced around $1,000, but development delays could push it to 2028; CEO John Ternus is pushing the robotics work forward. Source: Ars Technica.
Anthropic Residency: a paid, fixed-term research program opens, and no machine learning background is required. Anthropic opened applications for its paid, fixed-term research program for academic researchers, technical founders and experienced engineers who want to work on making AI safe and beneficial. No prior machine learning experience is required, and the deadline is November 22. Source: Anthropic (X).
Rumors
Sam Altman: the next GPT is coming “soon”, and 10,000 instances of it worked autonomously on a task for days. According to fan and leak-tracking accounts, Sam Altman said on Monday at the ADIPEC conference in Abu Dhabi that 10,000 agents of the upcoming next GPT model worked together for about four days without a management hierarchy; the reports link this model to the internal model working on the Navier–Stokes problem. Reliability: a second-hand summary of a public talk; the model's name and release timing are unofficial. Source: chetaslua (X), kimmonismus (X).
What this means for you: Monday showed that the price and limits of the fastest-growing AI services can change quickly: Grok Bot subscribers got tighter limits overnight, and ChatGPT and Codex will draw from a shared token allowance. If your team relies on these tools every day, it pays to know what each task really costs, and you can track that with our transparent, per-model pricing, while alongside Anthropic's warning about self-accelerating AI, our AI governance and risk management service helps your company use AI under rules agreed in advance.