Skip to Content

Latest AI news — November 20, 2026

New models: the Inkling 2 Mini looped model has been released, and Tencent has published the Hy4 flagship with open weights. Tools & features: ChatGPT Space is also reaching Plus subscribers, and Claude computer use is launching in the Team and Enterprise plans. Policy & safety: OpenAI has made it harder to steal models' reasoning, and Signal is blocking screenshots from AI agents. Business & society: an Anthropic mathematician claims to have proved the Riemann hypothesis with Claude, but Anthropic's share fell on its first full trading day. Rumors: Cognition is considering a fundraising at a $75 billion valuation, and Microsoft may be courting the Chinese customers Anthropic turned away.
November 21, 2026 by
Latest AI news — November 20, 2026
Krisztian Eyssen, Krisztian Eyssen

New models

  • Thinking Machines: Inkling 2 Mini has been released, a looped transformer that, according to the company, is small enough to run on a laptop but also strong enough for agent work. It runs the middle layers three times per token, so the model can keep thinking without writing the answer. It is a 26.4-billion-parameter mixture-of-experts (MoE) model, with 3.4 billion active parameters per token and, because of the looping, 5.4 billion weight activations; its memory is about 20 gigabytes. According to the company, it outperforms every open-weight model under 100 billion parameters on the DeepSWE v1.1 (48.2 percent) and Vals Index v2.2 (46.5 percent) tests, scored 20 points on the Artificial Analysis Intelligence Index, and is comparable to Qwen3.8 27B with roughly one-fifth as many active parameters. The weights are available on Tinker, Hugging Face and several local runtimes; the lab has also released a research variant with variable looping. Source: Thinking Machines, Artificial Analysis, Hugging Face.

  • Tencent: it has released the Hy4 flagship model with open weights, a native multimodal, 770-billion-parameter mixture-of-experts system of which 49 billion parameters are active per token, and which handles a one-million-token context. The cheaper Hy4 flash open weights already appeared in the previous roundup; this is the larger, flagship variant. Besides the weights, the company has also published a system card, with safety, cyber and biological risk assessments. Source: Tencent Hunyuan (X), Hugging Face.

  • Models in brief. humans& has opened two new test-time scaling (that is, more computation per answer) tiers of the Persimmon v0.1 chat model to everyone: according to the company, Ultra improved the result on its own eight-turn, multi-user Turing test from 18 to 48 percent, where 50 percent is already indistinguishable from a human. Apple has released the Depth Pro 2 model on Hugging Face, which estimates sharp, temporally consistent, meter-measured depth from a single camera image. Source: humans& (X), Hugging Face: Depth Pro 2.

Tools & features

  • OpenAI: ChatGPT Space, the interactive page and presentation editor, is reaching Plus subscribers aged 18 and over worldwide, and the company is starting to test dots, the proactive agents, with a narrow Plus group and with all ChatGPT Business Standard users. In Space, shared pages and slides can be created that may update themselves on the basis of connected applications; people on the Free plan cannot open the shared materials, and there is no public link. Space is taking the place of Library on the web, on the desktop and in the mobile app. Until now the dots were available mainly to Pro subscribers. The ChatGPT mobile app is also getting a new interface: the dot has its own tab, and Codex, Projects and Images can be pinned to the bar; this is launching first with Pro users aged 18 and over and with a narrow Plus group, excluding the European Economic Area, the United Kingdom, Switzerland and Korea. Source: ChatGPT release notes, Greg Brockman (X), Thibault Sottiaux (X).

  • Anthropic: computer use is launching in beta, gradually, in the Claude Team and Enterprise plans as well. With the feature, Claude sees the user's screen and can work in the permitted desktop applications; it is available on the new Claude Desktop interface, as well as on the web and in the mobile app if computer access is switched on. Claude Code version 2.1.330 has meanwhile received enterprise sign-in and a built-in diagram-drawing mode called easel. Source: Claude release notes, Claude Code changelog.

  • Google: Gemini can now also make GIFs, with the Gemini Omni 1.2 Flash video model. Moving images can be requested, from repeating photo effects to memes. According to TestingCatalog, the feature is for now available to Google AI subscribers in the United States, Japan, South Korea, Australia, Canada, the United Kingdom, Germany and France. Source: Gemini App (X), TestingCatalog (X).

  • Perplexity: the Computer agent also works directly in Chromium-based browsers through the Perplexity extension. The task can be handed over from the sidebar, and the agent works across tabs and tools. According to the CEO, setting up a teammate agent can now also be started from the input field, and the agent also uses the memory of the given conversation and of earlier ones. Source: Perplexity (X), Aravind Srinivas (X).

  • Hugging Face: it has shown that with a small adapter trained for a given agent harness (that is, the runtime around the model), the same model can be more accurate and cheaper. Qwen3.8-27B in the pi harness improved on Terminal-Bench 2.1 from 39 percent to 55 percent, while token use per task fell from 2,900 to 1,500. The lab trained a separate LoRA adapter for each of the pi, Hermes Agent and OpenCode harnesses, then also merged them into a shared adapter that can run in all three environments; it has published the recipes and the weights. Source: Hugging Face.

  • Developer and assistant tools in brief. In OpenAI's Responses API, the reasoning steps of the GPT-6.2 models are tied to the preceding conversation, so if an earlier message is edited, the API discards the reasoning after the change; according to the company this brings fewer tokens and more cache hits. Antigravity has received an official MongoDB Atlas (managed MCP) extension, with natural-language querying and vector search. According to Alexandr Wang, Meta's AI lead, the Muse agent is also available as a Chrome extension, from the browser sidebar. Source: OpenAI Developers (X), Antigravity (X), Alexandr Wang (X).

Policy & safety

  • OpenAI: in new research it has described how it defends against adversarial distillation, when someone trains, without permission, an unguarded copy from the model's outputs. The company especially wants to protect the protected chain of thought, because it can contain intermediate steps that are left out of the final answer. On an internal, adaptive test, the success rate of attacks on the updated GPT-6.2 Astra defenses fell from 54 to 0.2 percent compared with the earlier GPT-6 Astra defenses. OpenAI has this year also detected several coordinated distillation campaigns; the weekly interventions have affected thousands, and in some cases tens of thousands, of accounts, at a significantly accelerating pace over the summer, and they also share signs of the attacks with other model makers. Source: OpenAI, The Decoder.

  • Signal: by default it restricts screenshots in the mobile app so that AI agents built into operating systems, for example Gemini and Siri AI, cannot read private conversations. According to the organization, the agents of Google, Apple and Microsoft read data visible on the screen, and thereby undermine Signal's own privacy promise. They already took a similar step last year against the Windows 11 Recall feature. According to Signal, the new protection is not comprehensive. Source: Signal, The Verge.

  • US president Donald Trump said they would ban American companies from building data centers in other countries, and if they do so anyway, a “huge tax” would be imposed on them. At a White House press conference he said he had not told even his own staff this in advance. Alphabet, Amazon, Microsoft, Meta and Oracle have already built data centers outside the United States; the cloud providers did not immediately respond to CNBC's inquiry. Source: CNBC, The Verge.

  • Cursor: the coding platform owned by SpaceX told two congressional committees that it is considering phasing out open-weight models developed in China because of “geopolitical risk”. According to a report CNBC saw in advance, the Cursor Composer 2 model was built on the Chinese Moonshot AI Kimi K2.5 model. The report examined the US technology stack's reliance on Chinese models; Airbnb and DoorDash also conferred with the committees. Cursor, SpaceX and DoorDash did not respond to CNBC's inquiry. Source: CNBC.

  • A security researcher planted two million unique, traceable secrets (canaries) in AI services, and found that requests passing through routers (AI router, AI gateway) often leak. HD Moore, research director at Dragos, hid never-used AWS keys, passwords and unique addresses in individual requests. According to Ars Technica, there was a US hosting provider that visited a secret web address more than two months after it appeared in a single inference call; another came back days later as well. According to Moore, the safest option is a model run on one's own server, and the next is a direct contract with a front-line provider. Source: Ars Technica.

  • The US Senate has passed, by 62–33, the Bipartisan American Affordability and Jobs Act (BAAJA) proposal, which would speed up permitting for energy and technology investments, including data centers. The draft would ease some environmental rules and would reshape the regulation of electricity transmission; the Democratic Party was split on the vote. The proposal has not yet become law. Source: CNBC.

Business & society

  • Levent Alpöge, a mathematician at Anthropic, claims that together with Claude he has proved the 167-year-old Riemann hypothesis, the most famous unsolved problem in number theory, and has published the proof in the Lean formalization language. The hypothesis describes the hidden order in the distribution of prime numbers, and is among the Clay Mathematics Institute's Millennium Prize problems, each worth one million dollars. According to TechCrunch, several mathematicians, including Andrew Sutherland at MIT, have confirmed that the formalized proof compiles without error, which in Lean means that the machine has checked every step; the professional community is still working, however, on understanding the argument in human language, and it also has to be examined whether the hypothesis was formalized exactly. Alpöge has undertaken to provide an understandable explanation as well. According to TechCrunch, a human also announced a proof of their own on the same day; this too awaits independent review. Source: Scientific American, TechCrunch, The Verge.

  • Anthropic has closed its record IPO: with the full exercise of the greenshoe option it raised $93.5 billion, and together with the parallel private sale the gross proceeds are $119.5 billion. The company sold 1.127 billion Class A shares at $83 each; the further $12.2 billion came from the 147-million-share option, and among those taking part in the $26 billion private round were AMD, Google, Microsoft and Nvidia. After Thursday's 17 percent rise, on Friday, the first full trading day, the share fell 8.5–9 percent, according to The Guardian to $88.30. In the broader AI sell-off the Nasdaq Composite wiped out its 2026 gain and closed 5 percent below its level at the start of the year; according to the Financial Times the Nasdaq 100 lost 5.6 percent on Friday and 10.4 percent on the week. Nvidia's share fell 14 percent on Friday and closed down for a seventh week in a row, its longest such streak since 2008; according to CNBC, over that period it has lost 46 percent, about $2.6 trillion in market value. Source: Anthropic, CNBC: greenshoe, CNBC: share price, The Guardian, CNBC: Nvidia.

  • The evaluation organization METR has updated its much-cited time-horizon chart: some capabilities continue to improve roughly exponentially, but the models are weaker on tasks where success is harder to check and judgment is needed. The new task set also includes longer and more open-ended work, for example designing AI evaluations that METR staff could actually use. According to Technology Review, the best models already outperform almost everyone in software development, but here they did not find a single model-designed evaluation useful enough. They similarly lagged on a task that scores philosophical arguments. Source: MIT Technology Review, METR (X).

  • OpenAI: MedChemBench has been released, a 99-task test that measures how well a model can infer the next compound to examine from molecular structures, experimental results and trade-offs. GPT-6.2 Astra, at the highest reasoning level, scored 84.5 percent of the available points. The company separately showed how it is searching for new leads in the understanding and treatment of neurodegenerative diseases with researchers at Perturb and Scripps Research. Missed in our previous roundup: Isomorphic Labs, Demis Hassabis's AI-based drug-design company, has entered a strategic research collaboration with Merck (MSD); according to Hassabis they are jointly designing a new anticancer drug. Source: OpenAI, OpenAI (X), Isomorphic Labs, Demis Hassabis (X).

  • Meta: in its 2026 sustainability report it said it is reassessing the 2030 net-zero climate target and the near-term intensity goals, because demand for AI infrastructure has grown faster than they had expected. According to Ars Technica, the company's total greenhouse-gas emissions rose by almost 40 percent between 2024 and 2025, the largest annual increase among the four big US technology companies; Microsoft's was 25 percent, Google's 18 percent and Amazon's 16 percent. Source: Ars Technica.

  • According to a Bloomberg analysis of almost ten million Chinese job ads between July 2021 and June 2025, ads for occupations strongly exposed to artificial intelligence fell by about 10 percentage points more than those for less exposed roles. According to The Decoder, the effect is especially visible among those starting their careers and in creative professions, on an already weak labor market. Source: The Decoder.

  • Investments and companies in brief. Oracle said it has secured all the financing needed for its infrastructure contract with OpenAI, and that delivery does not depend on further supplier talks. According to Christine Lagarde, president of the European Central Bank, EU public money, including the EU budget, should bear the first loss on large chip-plant and data-center investments in order to attract private capital; in her view the plans for at most seven European AI “gigafactories” are not enough. Source: Oracle (X), CNBC.

Rumors

  • Cognition: according to Bloomberg sources, the developer of Devin is considering a fundraising of at least $2 billion, at a valuation of about $75 billion. The company has already raised more than $3 billion in 2026, including $2 billion in September at a $48 billion valuation. Reliability: medium–high (Bloomberg, anonymous sources; the company has not confirmed it). Source: Bloomberg.

  • According to The Information, Microsoft is courting the Chinese AI customers from whom Anthropic has withdrawn its service, from ByteDance to Alibaba. Reliability: medium (The Information, an exclusive report; the companies have not commented publicly). Source: The Information (X).

  • Nvidia: according to Reuters sources, the US securities regulator (SEC) is also examining the company's possible accounting problems. According to Reuters, part of the inquiry may be that Nvidia last year changed the geographic breakdown of revenue, and Singapore was dropped from the figures. According to Wired sources, what mainly troubled the board at the time of Jensen Huang's departure was not the national-security issue, but that Huang had not been honest with them for a year from October 2025. Reliability: medium–high (Reuters and Wired, anonymous sources; the company has not confirmed it). Source: Reuters, CNBC, Wired.

  • Anthropic: according to TestingCatalog, a new interface has started to appear in the Claude mobile app, with a bottom navigation bar and Chat, Projects, Code and Artifacts tabs. The changes mainly affect the navigation, the model selector and the attachment menu. Reliability: medium–high (an observed, gradual rollout; there is no official announcement). Source: TestingCatalog (X).

What this means for you: Claude computer use, ChatGPT Space and Perplexity's browser-based Computer show that agents are getting deeper access to your work tools, while Signal's screenshot protection, OpenAI's anti-distillation steps and the leaking AI routers are a reminder that well-scoped, logged agents are needed. The day around the Riemann hypothesis and the METR update also show that AI is already present in expert work, so it is worth reviewing which models and which data your company's tools can access, before a client or an auditor asks.

Latest AI news — November 20, 2026
Krisztian Eyssen, Krisztian Eyssen November 21, 2026
Share this post
Archive