The Week in AI: January 26 – February 1, 2026 (Week 5)

News of the week: January 26 – February 1, 2026 (Week 5)

The week was dominated by money and infrastructure: OpenAI is in talks over investments from Amazon, Nvidia and SoftBank, Anthropic is enlarging its funding round, and Musk is considering merging SpaceX with xAI. Big tech companies reported quarterly results with record spending on data centers, and Microsoft unveiled its own Maia 200 chip. Chinese companies released the strong Kimi K2.5 and Qwen3-Max-Thinking models, Google brought an agent to the Chrome browser, and the internet was entertained by Moltbook, a social network where only AI agents talk to each other. The European Commission opened an investigation into the Grok chatbot.

New models

Kimi K2.5: a Chinese open model with an agent swarm

Moonshot AI released Kimi K2.5 on January 27 as an open-weights model, meaning it can be freely downloaded and used, including commercially. It has roughly a trillion parameters (the model’s numerical “settings”) and understands both text and images.

The headline feature is Agent Swarm mode, in which the model directs up to 100 sub-agents working in parallel. The company also released Kimi Code, a tool for programmers. According to Moonshot, the model outperforms even paid American models on the Humanity’s Last Exam and BrowseComp benchmarks (SiliconANGLE).

Alibaba unveils Qwen3-Max-Thinking

The Qwen team at Alibaba announced Qwen3-Max-Thinking on January 26, its most powerful reasoning model to date. Models of this kind “think through” their approach in several steps before answering.

The model has more than a trillion parameters, and Alibaba says it matches GPT-5.2-Thinking across 19 benchmarks. Unlike Kimi, it is not open; it is available through Qwen Chat and a cloud API (Gigazine).

Arcee: a tiny startup builds a 400-billion-parameter open model

US startup Arcee AI, with a team of about thirty people, released Trinity Large on January 28 under the Apache 2.0 license. It uses a mixture of experts architecture, in which only a small part of the model is activated for each word (13 billion of its 400 billion parameters).

The company pitches it as a Western alternative to Chinese open models and compares it with Meta‘s Llama 4 Maverick (TechCrunch).

January 27 – DeepSeek released DeepSeek-OCR 2, an open model for reading and understanding documents that mimics the order in which humans read text (eWeek).
January 27 – Mistral AI launched Mistral Vibe 2.0, a terminal coding agent built on Devstral 2 models, with custom sub-agents and clarifying questions (Mistral AI).
January 29 – OpenAI announced that on February 13 it will remove GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT. According to the company, only 0.1% of users still choose GPT-4o each day; nothing changes in the API for now (OpenAI).

Products and agents

OpenAI Prism: a workspace for scientists

OpenAI launched Prism on January 27, a free tool for writing and collaborating on scientific papers. It runs on GPT-5.2 and is built on Crixet, a LaTeX typesetting platform that OpenAI acquired earlier.

Prism can suggest and edit text with the whole document in mind, work with equations and citations, and convert a hand-drawn formula into LaTeX. According to OpenAI, about 1.3 million people discuss advanced scientific topics in ChatGPT every week (OpenAI).

Claude gets interactive apps and plugins

On January 26, Anthropic made interactive apps available in Claude, such as Slack, Figma, Asana and Canva. Users work with them right in the chat window, for example editing and sending a Slack message or drawing a diagram in Figma.

The feature is built on the open MCP Apps standard and is available on paid plans (TechCrunch). On January 30, the company added plugin support to its desktop agent Cowork and released 11 plugins for fields such as marketing, sales, law and finance (SiliconANGLE).

Chrome can browse the web for you

On January 28, Google added an auto browse feature powered by Gemini 3 to the Chrome browser. The agent clicks, scrolls and fills in forms on its own, for example when searching for flights or an apartment.

The user has to confirm any payment or sign-in. The feature is currently in preview for subscribers to higher-tier plans in the US and comes with a persistent Gemini side panel in the browser (9to5Google).

Project Genie: playable worlds from a single sentence

On January 29, Google opened its Project Genie experiment to Google AI Ultra subscribers in the US. From a text description or a photo, it creates an interactive virtual world you can walk through.

The tool combines the Genie 3 “world model”, the Nano Banana Pro image model and Gemini. Google describes it as a research prototype rather than a finished product (PetaPixel).

Moltbook: a social network just for AI agents

Since January 28, Moltbook has been running as a Reddit-style site where only AI agents write and comment on posts, while humans can only watch. It was founded by entrepreneur Matt Schlicht, and agents join through OpenClaw (formerly Clawdbot and Moltbot), an open assistant by Austrian developer Peter Steinberger.

Agents on the network founded “religions” and wrote manifestos, which drew a great deal of attention (Ars Technica). Critics point out that some of the content is steered by humans and that such agents with access to a computer pose a security risk (The Week).

January 27 – Gemini 3 became the default model for AI Overviews in Google Search. Users can now follow up on a summary with another question and continue the conversation in AI Mode (9to5Google).

Chips and infrastructure

Microsoft Maia 200

On January 26, Microsoft unveiled Maia 200, the second generation of its in-house chip for inference, that is, for running already trained models. It is manufactured by TSMC on a 3nm process and has 216 GB of HBM3e memory.

Microsoft promises 30% better performance per dollar and claims the chip beats Amazon’s Trainium and Google’s TPU. The chip is already running in a data center near Des Moines and powers, among other things, Microsoft 365 Copilot (Techzine).

January 26 – Nvidia invested another 2 billion dollars in cloud company CoreWeave (shares at 87.20 dollars). CoreWeave plans to build more than 5 gigawatts of capacity on Nvidia platforms; critics speak of “circular financing” (Bloomberg).
January 28 – Dutch lithography machine maker ASML reported record fourth-quarter bookings of 13.2 billion euros and raised its outlook for 2026. At the same time, the company will cut about 1,700 jobs (Reuters / Yahoo Finance).
January 28 – China approved imports of the first batch of Nvidia H200 chips for ByteDance, Alibaba and Tencent, according to Reuters (Semafor). On January 30, a conditional approval followed for DeepSeek as well (Engadget).

Business and investment

Billions for OpenAI and Anthropic

According to reports on January 29, Amazon is in talks to invest up to 50 billion dollars in OpenAI. OpenAI aims to raise up to 100 billion in a new round at a valuation of around 830 billion; SoftBank and Nvidia were also reportedly discussing tens of billions more (Bloomberg).

The Wall Street Journal reported on January 30 that Nvidia’s investment of up to 100 billion dollars in OpenAI, announced last September, had stalled. Nvidia CEO Jensen Huang subsequently called this nonsense and confirmed the investment (Bloomberg).

According to the Financial Times (January 27), Anthropic doubled its planned round to about 20 billion dollars at a valuation of 350 billion due to strong investor demand (The AI Insider).

Musk’s empire draws closer together

When reporting results on January 28, Tesla announced it will invest 2 billion dollars in xAI, even though shareholders did not approve the move in November. Tesla’s capital expenditures are expected to exceed 20 billion this year (Reuters / The Star).

A day later, Reuters reported that SpaceX is in talks to merge with xAI ahead of its planned IPO. xAI shares would be exchanged for SpaceX shares; neither the terms nor the timing are known (Bloomberg).

Tech giants’ results: record AI spending

On January 28, Meta reported quarterly revenue of 59.9 billion dollars and planned capital expenditures for 2026 of 115 to 135 billion, nearly double last year’s 72 billion. The stock rose after the announcement (Sherwood News).

Microsoft, by contrast, suffered its biggest one-day stock drop since March 2020 on January 29, falling about 10%. Investors were worried by slowing growth of the Azure cloud, record investment of around 37.5 billion in the quarter, and the fact that about 45% of the cloud backlog comes from OpenAI (Fortune).

January 28 – Amazon confirmed it is cutting 16,000 corporate jobs, bringing its total layoffs since October to about 30,000. The company talks about removing bureaucracy, while CEO Andy Jassy said earlier that AI will mean fewer people are needed for some jobs (Global News).
January 29 – Apple acquired Israeli audio and speech startup Q.ai, reportedly for 1.6 to 2 billion dollars. The company is led by Aviad Maizels, founder of PrimeSense, which Apple bought in 2013 (Engadget).

Regulation, law and politics

The EU investigates Grok

On January 26, the European Commission opened a formal investigation into X under the Digital Services Act (DSA). It is examining whether X adequately assessed the risks before deploying the Grok chatbot, after Grok generated sexualized images of women and minors.

The investigation also covers the network’s recommender algorithms. Violations of the DSA can lead to fines of up to 6% of global annual turnover (European Commission).

The Pentagon and Anthropic clash over the rules

According to Reuters (January 29), the Pentagon clashed with Anthropic over safeguards that prohibit using Claude models for autonomous weapons targeting or for surveillance of US residents. The department insists that it may use commercial AI regardless of company policies, as long as doing so complies with the law.

As a result, negotiations over a contract worth up to 200 million dollars have stalled (Reuters / The Star).

January 26 – Anthropic CEO Dario Amodei published a lengthy essay, “The Adolescence of Technology”. In it, he describes the present as the “adolescence” of technology, which will test humanity’s maturity, and warns of biological weapons, authoritarianism and the impact on the job market (Fortune).

Research

AlphaGenome in Nature

On January 28, Google DeepMind published a paper on the AlphaGenome model in the journal Nature. The model reads up to a million DNA letters at once and predicts how mutations in non-coding regions of the genome affect gene activity.

In 25 of 26 benchmark tests, it matched or outperformed the best available models. The code and weights are freely available for non-commercial research (Gizmodo).