Skip to Content

Latest AI news — November 15, 2026

New models: Meituan has released LongCat-2.5, a 1.6-trillion-parameter agent model trained on Chinese hardware, with open weights. Tools & features: Claude Fable 5.2 paints in code, brushstroke by brushstroke; weekly hours worked by Perplexity Teammate agents have grown elevenfold. Policy & safety: Bessent is urging an outside CEO at Nvidia, Congress would tighten chip exports; the US and its allies would jointly protect AI supply chains from China. Business & society: Anthropic's IPO is seven times oversubscribed; Jensen Huang is considering a return. Rumors: according to Bloomberg, Apple is planning Siri-based smart glasses.
November 16, 2026 by
Latest AI news — November 15, 2026
Krisztian Eyssen, Krisztian Eyssen

New models

  • Meituan: the Chinese food-delivery company has made publicly downloadable the weights of the 1.6-trillion-parameter LongCat-2.5 model, which was trained entirely on Chinese hardware. The mixture-of-experts (MoE) model, meant for long, multi-step agent tasks, mainly programming, uses about 48 billion parameters per token, handles a one-million-token context and 128,000-token replies through the API, and understands text, images and video. The company has also released a 4-bit version optimized for Chinese chips; the standalone running guide assumes 194 Huawei Ascend accelerators and gives no instructions for Nvidia GPUs. On the Terminal Bench v2.1 command-line test it scored 88.4 points; in Meituan's table Claude Opus 5 stands at 89.1 and GPT-5.6 Sol at 88.8; the latest Opus 5.5 and GPT-6.1 Sol are not in the comparison. Source: The Decoder.

Tools & features

  • Anthropic: without an image-generation model, Claude Fable 5.2 "paints" with code: it assembles the picture brushstroke by brushstroke in a simulated painting engine. According to a demo by Anthropic staff member Jake Eaton, every image is a 1,000-to-2,000-line Python program that also simulates the texture of the brush hairs, the application of the paint and the painting-over of mistakes; a shopping cart standing in a parking garage, which looks like a photograph, was made from about 700,000 brushstrokes. It does not use an external image model, or a pre-rendered or traced source image, and for the styles the model worked out its own artist personas with invented backstories. Source: The Decoder.

  • Perplexity: according to CEO Aravind Srinivas, the number of weekly hours of work done by Teammate agents has grown elevenfold (a week earlier it was still sevenfold), and it is expanding by more than 50 percent a week. In a single week there was a more than threefold rise in the number of workers for whom Teammate completes at least one extra workday (8+ hours) or one extra workweek (40+ hours) of work. He added that Teammates can be built on any frontier model in Perplexity. Source: Aravind Srinivas (X), Aravind Srinivas (X).

  • Developer tools in brief. Simon Willison has released version 0.6a0 of Datasette Agent: the plugin that lets you ask questions about databases now offers an agent prompt field directly on the Datasette home page, where tables and queries can be referred to with a hash sign. A video reached the Hacker News front page in which the maker, with the help of Claude Code and Codex, reverse-engineered the engine control unit (ECU) of his 1994 Jeep through a $3 Arduino. Source: Simon Willison, YouTube.

Policy & safety

  • Nvidia affair: Treasury Secretary Scott Bessent is urging a "dramatic shake-up" at Nvidia, and in Congress both parties are pushing for tighter chip exports. According to Bessent, the company should choose a CEO from outside, and within ten days he wants an answer as to whether Nvidia pushed the "apparently prohibited" chip sales to China in order to increase its revenue. According to The Verge, a bill on the security of remote access that has been stalled since 2025 (the Remote Access Security Act) is back on the agenda, and House Speaker Mike Johnson spoke of a "growing consensus" on restricting chip exports to China. According to Bloomberg, a House committee chair said Jensen Huang has agreed to testify, but there is not yet a date; according to The Information, the Commerce Department is preparing to close the loophole through which Chinese AI firms reach restricted Nvidia chips via data centers overseas. Source: Bloomberg, CNBC, The Verge, Bloomberg: hearing, The Information (X).

  • Megaspeed: the Singapore cloud provider has opened an internal investigation, after the Wall Street Journal linked it to the brokering of some $12 billion of Nvidia chip orders to China. According to the company, Alice Huang left on August 30, 2025, and today no longer has a role at the firm; according to documents seen by the WSJ, between March 2024 and June 2025 Megaspeed ordered hundreds of thousands of export-restricted chips for Chinese buyers. According to The New York Times, Commerce Secretary Howard Lutnick pushed for an easing of chip rules that affected a client of his former firm, Cantor Fitzgerald, and according to Bloomberg, Nvidia turned to the Taiwanese authorities after one of its managers was accused of chip smuggling, but TSMC did not want to intervene. Source: CNBC, The Decoder, New York Times, Bloomberg.

  • Pax Silica: the United States and its allies, including Japan, the United Kingdom and India, have agreed to draw up a joint list of the AI technologies and supply chains to be protected from China. According to the Financial Times, the parties have signed the agreement on this. Source: Financial Times, Bloomberg.

  • South Korea: the candidate who achieved the best result in the tax-advisor exam has admitted cheating with generative AI. The candidate also came first in the April and July rounds, and in the first of them got only one of 160 questions wrong; according to his account he photographed the questions with a microcamera, sent them to an AI tool, and looked at the answers on a smartphone. The case is also putting pressure on the organizers ahead of Thursday's national university entrance exam. Source: The Guardian.

Business & society

  • Anthropic: according to Bloomberg's sources, the IPO is already about seven times oversubscribed, and together with a separate private placement the company and its investors may raise as much as $90.4 billion. The oversubscription, previously described as sixfold, has grown further. According to CNBC, the Nasdaq listing is expected on Thursday, at a share price of $70–78; AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC are together investing $26 billion at the IPO price, and Anthropic's prospectus contains more than $900 billion in unpaid spending commitments. Source: Bloomberg, CNBC.

  • Nvidia: according to Financial Times sources, Jensen Huang has discussed a return to the CEO post with confidants, and his supporters have set up a "war room" to secure White House support for this as well. According to The Wall Street Journal, Huang himself had already been pulling back Nvidia's aggressive customer financing, and had asked emerging cloud providers to collect customer payments in advance; interim CEO Colette Kress may go even further. According to The Information, AMD and Broadcom would not be able to take over the large credit guarantees that Nvidia gave to cloud and data-center developers. Source: Financial Times, Wall Street Journal, The Information (X).

  • OpenAI: it has launched its science blog, which publishes first-hand accounts by researchers of how they use AI for discoveries. In the first post, physicist Nathaniel Craig shows how GPT-6 Astra helped shape a possible non-perturbative lattice definition of the Standard Model of particle physics into a detailed proof. (The post appeared late on Sunday evening, at 11 p.m.) Source: OpenAI, OpenAI: Standard Model.

  • Samsung: according to the Financial Times, the South Korean giant would offer the building, operation and energy supply of AI data centers from a single hand. According to the FT's sources, Samsung's IT subsidiary is in early talks about Anthropic and OpenAI moving into its data centers as tenants. Source: Financial Times.

  • Stock market: according to the Financial Times, the shares of companies tied to the building of data centers have lost a total of $4.2 trillion in value, as the AI-driven market boom appears to be running out of steam. According to CNBC, the industry is also under pressure from rising interest rates and the fall of Nvidia's leader, so Anthropic's Thursday IPO is the next big test of the AI market. Source: Financial Times, CNBC.

  • Australia: according to an analysis by the Reserve Bank of Australia (RBA), AI has so far barely eliminated jobs, but in the occupations most affected, entry-level workers are already faring worse. The number of job ads has fallen in every group since ChatGPT appeared in 2022, most steeply in programming, office work and customer service; there, employment among 15-to-24-year-olds and recent graduates is weaker, and unemployment higher, than among older workers. The author cautions against hasty conclusions: the study does not prove a causal link, and only measures theoretical exposure. Source: The Decoder.

Rumors

  • Apple: according to Bloomberg's sources, Apple would unveil its smart glasses at WWDC and would ship them by the end of 2027, with the focus on Siri AI integrations for third-party iPhone apps. According to the report, Apple now treats AI as a more serious priority. Reliability: high (Bloomberg, internal sources; Apple has not confirmed it). Source: Bloomberg.

  • Google and OpenAI: according to a leaker, the first checkpoint of the previously mentioned Google "Krypton" model has appeared, under the name "krypton-a", and "Sol 6.2" is also on the way. According to the leaker, Krypton is outstanding at research tasks, and Sol 6.2 is very frugal with tokens; neither company has confirmed it. Reliability: low–medium (a single leaker, with no official confirmation). Source: Chetaslua (X).

  • Google: according to TestingCatalog, the settings page for the forthcoming "Comic" artifact in Gemini Notebook has been updated, and style examples are already visible on it. According to the leaker, this suggests that the release of the Nano Banana 3 image model may be close. Reliability: medium (TestingCatalog, based on traces found in the code, with no official confirmation). Source: TestingCatalog (X).

What this means for you: the scandal around Nvidia and the tightening export rules show that depending on a single chip or model vendor is a business risk, so it is worth deciding deliberately on vendor choice and substitution options, before a supplier's situation suddenly changes. If you also want to rely on more than one model in everyday work, eYssen AI Gate reaches Claude and GPT with a single API key, routes requests to the level that fits the task so simpler work does not pay top-model prices, and automatically fails over if a provider goes down. The Perplexity Teammate data also show that AI agents are already replacing real hours of work, so it is now worth assessing which of your recurring tasks can be handed over to them.

Latest AI news — November 15, 2026
Krisztian Eyssen, Krisztian Eyssen November 16, 2026
Share this post
Archive