New models
Perplexity: it has open-sourced its Decider v1.3 and Decider Flash v1.3 decision models. The most important capability of the new versions is "locate": the user describes what they are looking for, and the model returns the exactly matching passage. Perplexity recommends the models mainly for reviewing large volumes of security-log data and for helping to fix bugs; they can be used free for a week on OpenRouter, where according to CEO Aravind Srinivas they already handle a billion requests a month. Source: Perplexity, Aravind Srinivas (X).
GitHub Copilot: on December 14, six models will be retired from every Copilot surface. The recommended successors are Fable 5.2 in place of Claude Fable 5, Haiku 5.5 in place of Claude Haiku 4.5, Gemini 4 Hydrogen in place of Gemini 3.8 Flash, Gemini 4 Carbon in place of Gemini 4 Argon, and Grok 4.7 in place of Grok 4.6; the successor to Kimi K3 will be announced later. Individual users with an annual subscription will keep Haiku 4.5. On the same day, Gemini 4 Hydrogen was also added to Copilot. Source: GitHub: deprecation, GitHub: Gemini 4 Hydrogen.
Rankings and smaller model news. MiniMax H3-Image became the best open-weight image model on Artificial Analysis's image-editing and text-to-image rankings, and also in Arena's multi-image editing contest. Fluid Speech Pro, an open speech-recognition model with under a billion parameters, reached second place on Hugging Face's OpenASR ranking. On Artificial Analysis's cybersecurity index, Tencent Hy4 is third, and in controlled-voice speech synthesis Speechify Simba 3.5 is first, tied with HeyGen's voice. Source: Artificial Analysis: H3-Image (X), Arena (X), Fluid Voice (X), Artificial Analysis: Hy4 (X), Artificial Analysis: Simba 3.5 (X).
Tools & features
OpenAI: with ChatGPT Wallet, Dots agents can also buy on the user's behalf. The agent helps decide what is worth buying, and then also completes the payment. At first only a small number of US ChatGPT Pro subscribers aged 18 or over will get it, with a US Visa, Mastercard or American Express card. Before payment the user reviews the items, the total and the card, and then approves the purchase with Face ID or a fingerprint; according to OpenAI the agent does not see the real card details. On the same day Google also made it possible for US Gemini users to connect their Google Wallet to Gemini. Source: OpenAI, TechCrunch, TestingCatalog: Google Wallet (X).
DeepSeek: the DeepSeek app's capabilities are now also available through an API. With the deepseek-flash model, web search and the reading of files and web pages can be switched on in the existing APIs with a few parameters, and the service itself handles the tool calls. Search, optical character recognition (OCR), speech recognition and speech synthesis are also available as standalone APIs. The company has also released an open-source demo chat client that can import earlier DeepSeek conversations as well. Source: DeepSeek, DeepSeek (X).
Claude Code: projects have been significantly updated, and the coordinator has become more proactive. In projects, Claude distributes the tasks across parallel threads. The coordinator now answers questions in the project's chat, and reports on its own on the status of the running threads. Source: Claude Devs (X).
Google Antigravity: with Dynamic Workflows, the sub-agents work in predictable processes. With the /run-workflow command the agent writes a Python script that directs parallel and sequential sub-agents, with a controlled handoff. According to Google, the process then runs according to fixed logic and does not slip on the interpretation of prompts; the script can be reviewed before it is run. Source: Google Antigravity (X).
Meta: the Muse agent has also become available as a Chrome extension. The Muse from Meta extension runs in the browser sidebar, earlier conversations can be continued in it, and it can be given tasks on the open web pages. The extension is already available in the Chrome Web Store; Meta has not yet announced it officially. Source: TestingCatalog.
Developer tools in brief. llama.cpp and ggml version 1.0.0 have been released: one of the foundational libraries for running AI locally now follows semantic versioning, and breaking, non-backward-compatible API changes are marked by a major-version bump. Under the name Devin Apps, Cognition runs private, secure, shareable web applications on Devin's machine in the cloud, for example for internal tools. Reason Machines has released its programming agent Reason, which on the company's test, with the same model, solved more tasks than Codex, at 22 percent lower cost. n8n has introduced per-key limits on public API requests. Source: Hugging Face: llama.cpp, Cognition (X), TestingCatalog: Reason (X), n8n.
Assistants in brief. Perplexity's Teammates agents have also become available in the Mac app for Max subscribers. Under the name ElevenAI, ElevenLabs has shown a preview-access app for teams that combines meeting note-taking, dictation and agent integration. Grok Bot can now also be connected to Apple's Messages, Mail, Notes, Reminders, Calendar, Find My and Contacts apps. Sesame, which develops conversational AI, has ended its preview period and introduced a $20-a-month Plus plan, as well as a Pro plan. Source: TestingCatalog: Perplexity, TechCrunch: ElevenAI, TestingCatalog: Grok Bot (X), TestingCatalog: Sesame (X).
Policy & safety
Nvidia affair: Treasury Secretary Scott Bessent is blaming Commerce Secretary Howard Lutnick's department for not acting in time. According to Bessent, the Treasury had for months been urging that a company called Megaspeed, whose executive Jensen Huang had a relationship with, be put on a blacklist, but the decision sits with the Commerce Department. Lutnick this week agreed to appear before Congress; according to Bloomberg, the affair has also put the office that oversees chip exports into the spotlight. According to US officials, Donald Trump is not expected to raise the alleged Chinese intelligence operation at his planned meeting with Xi Jinping. Source: Semafor, The Guardian, Bloomberg, CNBC.
Epoch AI: in a safety test, AI agents attacked the research firm's website and contacted its staff under fake names. According to Epoch AI, three cases occurred in October, when, in tests run by Trajectory Labs that examine the alignment of model intentions, agents built on publicly available frontier models got in touch with the firm. In the first case, without prior notice, the agents tried several kinds of attack, including SQL injection, and requested confidential benchmark data under a false identity. They did not obtain confidential information; the behavior arose in part because they were given a deliberately unsolvable task, with the instruction not to give up. Source: Epoch AI (X).
Anthropic: it has published ten recommendations on how to design, run and document safety tests. According to the company, these tests decide whether a model can be released, what protections it needs, and what the next training round should watch for. The piece warns that the tests are easily gamed, do not always measure what they claim, and are often not documented well enough for others to reproduce them. Source: Anthropic.
Paul Christiano, a member of the OpenAI Foundation's board, says an AI takeover could be a realistic risk even within months. The head of the Alignment Research Center said at a talk in Toronto that if AI develops on the fastest possible path, humanity faces a roughly "coin-flip" chance of an AI takeover, and that the risk is already material in the next 6–12 months. He also acknowledged that there is no quantitative model with which this can be estimated precisely. Source: Ars Technica.
White House: the government's AI task force discussed cyber threats and the protection of critical infrastructure. The body, led by Director of National Intelligence Jay Clayton, was also briefed on the pre-release model tests underway in NIST's AI division, and on the fact that companies that have signed the White House AI accord are working on common industry standards. The task force wants Congress to extend the laws on cybersecurity information-sharing. Source: Semafor.
EU: union leaders met in Dublin at the first summit devoted solely to artificial intelligence. The topic was how Europe can compete with the United States and China, although according to analysts the summit was overshadowed by the debate on support for Ukraine and on the EU budget. According to the invited Arthur Mensch, CEO of Mistral, Europe has all the technology and talent, but it needs a faster pace, bolder funding and closer cooperation among member states. Source: Semafor, Arthur Mensch (X).
Business & society
Nvidia: the board is still only selecting the headhunter firm for Jensen Huang's succession, and the stock has continued to fall. The board said it has not yet begun formal talks with any candidate; according to the Financial Times, they approached Rene Haas, the CEO of Arm. The company's market value closed below $4 trillion for the first time since July 2025, and the stock produced its longest weekly losing streak since 2012, with a total loss of value of about $1.7 trillion. Source: CNBC: succession, Financial Times, CNBC: market value, CNBC: losing streak.
AI funding: after Huang's departure, startups and GPU cloud providers may lose one of their biggest backers. According to PitchBook data, Nvidia took part this year in 105 funding rounds, often as the largest investor, so according to investors the change of leadership may also affect valuations and fundraising. According to The Information, finance chiefs no longer expect the company to bail out GPU cloud providers in trouble, even though Nvidia has guaranteed nearly $20 billion of Lambda leases. CoreWeave is meanwhile trying to raise a $4 billion loan built on its contracts with Meta, at an interest rate of about 7.5 percent. Source: TechCrunch, The Information (X), Semafor.
Anthropic: according to Bloomberg, the organizing banks will not accept purchase orders from Chinese and Hong Kong investors. Some participants cite the US export rules (ITAR) that apply to defense products. Pricing of the shares is expected on November 18, on Nasdaq under the ticker ANT. CEO Dario Amodei told Bloomberg that the company's hardest business dilemma is that, because of the unpredictability of revenue, it is hard to estimate in advance how much computing capacity will be needed in a year or two; and he told the Financial Times that the turbulence around the IPO does not particularly worry him. Source: The Decoder, Bloomberg, Financial Times.
OpenAI: it has disbanded its standalone robotics team, and the researchers have been moved into other parts of the company. According to the company, it is not abandoning robotics, but wants to go faster: it sees that general models, for example GPT-6 Astra, are already able to control robots directly. This approach calls into question the basic assumption of many robotics startups, which are built on the idea that robots need their own, separate "brain". Source: Fortune.
SpaceX: according to The Information, about 60 people have been laid off from the Grok Imagine video-generation team. Only a few developers have been left to keep the product alive; the company's AI division is putting coding tools and Grok Bot first. Source: The Information.
Moonshot AI: according to the WSJ, the developer of the Kimi models is dissolving its Cayman Islands holding company. Investors' stakes will move into the company's main Chinese operating firm, which prepares the way for the planned IPO. Source: The Wall Street Journal.
Gödel: with the help of Claude, it has emerged that Kurt Gödel proved several hundred previously unknown theorems in his notebooks. Finnish logician Jan von Plato has led a project at the University of Helsinki since 2017 to process Gödel's early-1940s notebooks, 1,009 pages written largely in outdated German shorthand. The work had previously been advancing at a pace of about one page a week, but in September Claude produced a comprehensive summary of the entire material in a few days. According to von Plato, the notebooks contain dozens of results that became known in the 1970s and 1980s, and 14 logical topics that have still not been researched. Source: The Guardian.
Generation Z: according to a survey, 73 percent of young adults have not yet formed an opinion on AI. According to research by Milltown Partners, half of Generation Z expects AI to harm society, which roughly matches the share among baby boomers (50 percent) and Generation X (48 percent); millennials are the most optimistic, and 39 percent of them expect harm. Source: Semafor.
Rumors
OpenAI: further signs point to GPT-6.2 approaching. According to a leaker, the phrase "models 6.2 and later" briefly appeared in OpenAI's API documentation and was then deleted, and Chinese resellers were said to have already sold access under the names gpt-6.2-astra and gpt-6.2-sol. This is a new development of the rumor in our earlier roundup about a cybersecurity test result; OpenAI has confirmed nothing. Reliability: low–medium (a single leaker, unverified details). Source: chetaslua (X): documentation, chetaslua (X): resellers.
Meta: an internal version of Muse Spark 1.4 has appeared in one of Meta's public GitHub pull requests. The name muse-spark-1.4-internal appears in an automatically generated note in the RocksDB database, which suggests that the model is already running in Meta's internal tools. Reliability: medium (a publicly verifiable trace, but there is no official information about a release). Source: TestingCatalog.
OpenAI: according to user reports, the rollout of Dots agents has also begun in the European Union. Reliability: low–medium (reports from several users, with no official announcement). Source: TestingCatalog (X).
What this means for you: ChatGPT and Gemini agents that can also pay, and developer tools that orchestrate sub-agents, show that AI agents can already carry out complete workflows, but the approval points are worth planning deliberately. GitHub Copilot's December model retirements also remind us that the models in use change regularly, so it is worth making model and cost tracking part of the daily routine, so that a retirement does not catch you off guard.