New models
Rankings: Claude Fable 5.2 and Claude Opus 5.5 are tied at the top of Artificial Analysis's refreshed coding-agent index, and Gemini 4 Carbon took first place in the MathArena mathematics contest. Version 2.1 of the index measures the models in an agent environment, and the FrontierCode test (2.0 instead of 1.1) and Terminal-Bench have been updated in it; the two Claude models stand at 58 points. On MathArena, Gemini 4 Carbon pushed GPT-6.1 Sol out of first place, with Fable 5.2 second. The price, however, is substantial: with the thinking settings used in the test, Carbon costs about 15 times as much as GPT-6.1 Sol, and Fable 5.2 about 40 times as much, and Carbon's most common error is exceeding the time limit. On the decision-model index, Subquadratic SubQ Snap 1.2 improved by five points and is tied for the lead. Source: Artificial Analysis: coding-agent index (X), The Decoder: MathArena, Artificial Analysis: SubQ Snap 1.2 (X).
Tools & features
Anthropic: custom themes can now be created in the Claude Code desktop app. With "mods" the colors can be changed, and an animated background can also be set; it is enough to describe to Claude how the interface should look, or to start from the samples in the theme-examples plugin. Source: Claude Devs (X).
Google: Google Research has presented what they achieved with Antigravity's /teamwork multi-agent system. Google and its university partners, with agent teams built on the Gemini 4 Argon, Carbon and Hydrogen models, have been working on problems that have been open for years, from information theory through formal verification and molecular structures to kernel optimization, always together with human experts. The system tries solutions in parallel, and automatic checkers filter them (for example the Lean 4 proof checker, or unchanged test suites); each agent only gets the files and tools needed for that step, so that detail is not lost in work that lasts several days. The tools are available to anyone in Antigravity with the /teamwork and /boost commands. (Missed in our previous roundup:) Source: Google Research.
Anthropic: old prompting habits are holding Claude Opus 5.5 back, so the company has issued new usage advice. Phrases such as "think thoroughly" can be dropped, because the model adjusts its thinking on its own to the task and to the chosen effort level; without them the replies started faster, with no drop in quality. Anthropic recommends starting with the "Medium" level: according to its internal tests, Opus 5.5 then matches the "High" level of its predecessor, Opus 5. Anyone who regularly runs Sonnet at the highest levels pays roughly as much as for Opus 5.5, and it is worth describing the goal, the form of the output and the conditions for follow-up questions in a single message. Source: The Decoder.
In brief: Seedance 2.5 Flash in HeyGen, updated Qwen3.8-27B quantizations from Unsloth, and Musk has clarified Grok's text watermark. According to HeyGen, ByteDance's faster video model, Seedance 2.5 Flash, is up to twice as fast as the full Seedance 2.5, while its quality stays close to it. With the Dynamic v3.1 update, Unsloth has remade the compressed (quantized) versions of Qwen3.8-27B that can be run locally. According to Elon Musk, the watermarking of Grok's text only works where the law requires it, or where the user asks for it; the previous evening it was also turned off by default in the SpaceXSI API. Source: HeyGen (X), Unsloth, Elon Musk (X).
Policy & safety
Cursor: the October security incident began with an access key that leaked at a supplier, Fireworks. After a two-week investigation the company said the attacker, using a key Cursor had issued to Fireworks, copied coding-agent logs from a read-only store, mainly from 2025, from sessions in which privacy mode was not switched on. With the credentials obtained this way, on October 29 the attacker placed malicious copies of the Claude Code extension on Cursor's extension marketplace. The logs of less than 0.05 percent of users were exposed, and those users were notified individually; the attacker's last activity was blocked on November 1. Cursor has taken all inference traffic off Fireworks, has replaced every internal credential, and has rebuilt the extension marketplace. Source: Cursor.
Italy: the data-protection authority has fined Meta €805 million because the Facebook Android app had access to users' photos. The "gallery suggestions" feature was on by default, and with a single, vaguely worded permission it could read the photos, videos and metadata stored on the phone; according to the authority this may have affected some 20 million Italian users. Meta has 60 days to turn the feature off by default and to ask for separate consent; the company is appealing, and considers the fine disproportionate. The proceedings did not examine whether the content was used for AI training (which Meta denies), but the authority is separately investigating the feature launched in 2026 that processes gallery content in the cloud. Source: Garante Privacy, CNBC.
Research: in many web apps built with Lovable, personal data are reachable without signing in. Germany's Crissy Field examined 159,041 Lovable projects and in many cases found unprotected personal and confidential data, more than 200,000 chat messages in total; at one provider more than a thousand records containing passport numbers and citizenship were accessible. According to the researchers the fault lies mainly in the database-access settings, and Lovable's terms of use shift the responsibility onto the customers. Handelsblatt published the findings in German in early November, and the English version reached the Hacker News front page on Saturday. Source: Crissy Field.
United States: the Commerce Department is preparing to close the loophole through which Chinese AI firms reach restricted Nvidia chips in data centers abroad. According to The Information, the regulation targets leasing through overseas data centers; after the Nvidia scandal, pressure on the office that oversees the export restrictions had already been rising in the preceding days. Source: The Information (X).
Jacob Coxon: Anthropic could already retrain the Mythos model with a fraction of the original compute. According to the former OpenAI and Anthropic researcher, one chart in Anthropic's IPO presentation shows that the efficiency of the algorithms is improving ever faster, and that is part of why he left after barely four months. In his view, algorithmic improvement will soon be automatable, so next year he would pause the development of frontier models so that competitors can catch up. Source: The Decoder.
Australia: Liberal senator Andrew McLachlan says artists' copyright must not be sacrificed for AI investment. The Labor government's AI package also covers data centers, online safety and an amendment to copyright law; according to McLachlan, the building of data centers must not take priority over environmental protection either, and he raised the possibility that weakening copyright could be an unconstitutional taking of property. Source: The Guardian.
Business & society
Nvidia: more than 115,000 people have signed the petition asking for Jensen Huang to be reinstated, and two large investors are urging a cautious succession. According to the petition site the signatories include 5,965 current Nvidia employees, as well as Elon Musk, Amjad Masad (Replit), Tim Sweeney (Epic Games) and Garry Tan (Y Combinator). In a letter, Brad Gerstner (Altimeter) and Gavin Baker (Atreides) asked the board not to rush the succession, and also to consider Huang's possible return. According to Musk, many of Huang's key supplier agreements were based only on a handshake, so some of them may be lost without him. Source: TechCrunch, CNBC, Elon Musk (X).
Huang's departure is also being felt in Asia and in the world of open models. Several Asian governments have reaffirmed chip-delivery commitments, while others are watching the impact on local projects; according to Reuters, the future of more than $3 billion of Chinese compute projects built on restricted Nvidia chips is uncertain. According to Nathan Lambert, Huang's resignation is a major setback for US open models, because the grants he distributed are very likely to dry up. Source: TechCrunch, Interconnects.
DeepSeek: according to The Information, Tencent and CATL have also joined the funding round at a $92 billion valuation. The investors also include firms that have been described as one of the main sources of black-market Nvidia chips. Source: The Information (X).
Anthropic: ahead of next week's IPO, more and more analyses are turning to the company's value. According to The Information, Anthropic will go public at a valuation of at most $1.4 trillion, while one analyst estimates $1.7 trillion, and $3.4 trillion in the most optimistic scenario. The Financial Times calculated that Sam Bankman-Fried's $500 million investment in 2021 would be worth about $37 billion at a valuation of some $1.5 trillion; the stake was sold in 2024 for about $1.3 billion. Source: The Information (X), Financial Times.
Google: according to The Information, the company is looking at acquiring AI-based ad-creation tools. Google is also paying the Hightouch startup to move advertisers onto its customer-data tools; the aim is to make the ads that appear in AI features more useful, because advertisers complain that they cannot see how they perform. Source: The Information (X).
Databento: the market-data company has received a $100 million Series C at a $1.1 billion valuation. The 39-person firm has enterprise contracts with five of the six largest frontier-model AI labs, and its revenue has more than doubled since this year's Series B round. Source: FinTech Global.
Tesla: selected customers can also preorder the driverless Cybercab for private use. Alongside the $29,990 price comes a mandatory $299-a-month self-driving subscription, and the car may only travel in zones designated by Tesla, which can be changed at any time, in a few Texas cities. According to Elon Musk the monthly fee is too high, and about $115 would be realistic. Source: TechCrunch.
Foundation: the military humanoid-robot startup also backed by Eric Trump has been unable to pay its staff. According to CNBC's sources, the company missed payroll on October 30, while its CEO, Sankaet Pathak, last year borrowed $4.6 million from the company to buy a luxury house. Source: CNBC.
Research: generative AI uses 1.4 to 180 times more energy for text that is not in English. Researchers at the University of Michigan measured this on three open models: most Western European languages need at most twice the energy, Hindi 7.9 times as much, and Bengali 15.3 times as much, because for these the models produce more tokens (text units), and the answers are often less accurate too. According to their estimate, the native languages of some 5.2 billion people need at least twice as much energy as English. The study, published in early November, reached the Hacker News front page on Saturday. Source: University of Michigan.
Rumors
Google: according to a leaker, in the Arena contest, requests to Gemini 4 Hydrogen are being routed to the previously surfaced "Krypton" model. According to the leaker, their own tracking program detected this, and Krypton is strikingly better at motion graphics made purely from code; Google has not confirmed it. Reliability: low–medium (a single leaker, with no official confirmation). Source: Chetaslua (X).
What this means for you: the Cursor incident and the data leaks in Lovable apps show that an AI tool's security also depends on supplier keys and well-configured database permissions, so it is worth putting in place oversight and access rules in good time for the AI tools used by the team. The advice on Opus 5.5 is also a reminder that model choice and settings affect costs directly, so a regular review of model prices and effort levels can bring real savings.