The Week in AI: March 30 – April 5, 2026 (Week 14)

This week’s news: March 30 – April 5, 2026 (Week 14)

Money dominated the week: OpenAI closed a record $122 billion funding round and bought the tech talk show TBPN, Microsoft announced $10 billion for Japan, and Mistral took out a loan for its own data center. Google released the open Gemma 4 models under a permissive license, Microsoft unveiled three in-house MAI models, and Alibaba launched Qwen3.6-Plus for agents. The second big theme was software supply chain security: the leak of the Claude Code source code, the compromised axios package, and an attack via the LiteLLM library. Copyright and privacy lawsuits continued, and Hollywood screenwriters negotiated a new contract with rules for AI.

New Models

Google Releases Gemma 4 Under the Apache 2.0 License

On April 2, Google released the Gemma 4 family of open models, built on the technology behind the Gemini 3 models. It comes in four sizes: the small E2B and E4B variants for phones and offline devices, a 26B Mixture of Experts model (a model made up of multiple “experts,” only some of which are activated for each query), and a dense 31B model.

The biggest change is the Apache 2.0 license, which, unlike the earlier custom Gemma license, imposes no special usage restrictions. The model weights can be downloaded from Hugging Face, Kaggle, and Ollama; however, neither the training data nor the code has been published, which is why they are referred to as “open-weight” models. According to the Arena AI leaderboard, the 31B version ranks among the top three open text models. (Decrypt)

Microsoft Unveils Three In-House MAI Models

On April 2, the Microsoft AI division led by Mustafa Suleyman made three models developed in-house available, putting it in direct competition with its own partner OpenAI. They are available on the Microsoft Foundry platform and in the MAI Playground testing environment. (TechCrunch)

MAI-Transcribe-1 converts speech to text in 25 languages and, according to Microsoft, is more than twice as fast as the existing Azure Fast service.
MAI-Voice-1 generates speech, producing a minute of audio in a single second, and can also create custom voices.
MAI-Image-2 is the second generation of the image model, which first appeared in the testing environment back on March 19.

Alibaba: Qwen3.6-Plus Targets Agents and Coding

On April 2, China’s Alibaba introduced Qwen3.6-Plus, the first model in the new Qwen3.6 series. The model is available as a cloud service, has a context window (the amount of text a model can process at once) of 1 million tokens, and the company is aiming it at agentic coding and enterprise workflows. (Alibaba Cloud, CnTechPost)

Products and Agents

Cursor 3 Puts Agents at the Center of the Editor

The popular code editor Cursor released Cursor 3 on April 2. The former Composer side panel has been replaced by an agents window, where multiple AI agents can run in parallel, either locally, in isolated cloud environments, or on remote servers over SSH.

The agents are powered by the company’s own Composer 2 model, released in mid-March. Some developers welcome the change, while others worry that the editor is abandoning the “developer in charge, agent assists” principle. (SiliconANGLE)

Business and Investment

OpenAI Closes a $122 Billion Round

On March 31, OpenAI announced the close of a funding round worth $122 billion at a post-money valuation of $852 billion. It is the largest private round in Silicon Valley history; the company had originally talked about $110 billion.

According to reports, the largest contributions came from Amazon ($50 billion), Nvidia ($30 billion), and SoftBank ($30 billion). For the first time, retail investors were also able to take part through banks, contributing roughly $3 billion. OpenAI reports revenue of around $2 billion a month, and investors link the round to a possible stock market listing. (TechCrunch, The Next Web)

OpenAI Buys the Talk Show TBPN

An unexpected acquisition followed on April 2: OpenAI bought TBPN, a daily tech talk show popular in Silicon Valley. It is the company’s first media acquisition; the show will fall under the strategy department led by Chris Lehane.

OpenAI promises that TBPN will keep its editorial independence in choosing guests and topics. The price was not disclosed; according to unofficial sources, it was in the “low hundreds of millions of dollars.” (OpenAI, Bloomberg)

Microsoft Invests $10 Billion in Japan

On April 3, Microsoft announced that it will invest 1.6 trillion yen, or about $10 billion, in Japan between 2026 and 2029. The money will go toward AI computing capacity in partnership with SoftBank and Sakura Internet, toward cybersecurity, and toward training one million engineers and developers by 2030. The plan was presented by company president Brad Smith during a visit to Tokyo. (Reuters / Yahoo Finance)

Mistral Borrows $830 Million for a Data Center

On March 30, France’s Mistral AI announced its first debt financing: $830 million from a group of banks including Bpifrance, BNP Paribas, and HSBC. It will use the money for a data center in Bruyères-le-Châtel near Paris with 13,800 Nvidia GB300 chips and 44 megawatts of capacity. By the end of 2027, it aims to have 200 megawatts of capacity in Europe. (The Next Web)

Other Investments and Company News

March 30 – Qodo, a tool for reviewing AI-written code, raised $70 million. (TechCrunch)
March 30 – ScaleOps, which saves computing capacity in Kubernetes environments, raised $130 million at a valuation of $800 million. (TechCrunch)
March 30 – Starcloud, which wants to build data centers in orbit, raised $170 million. (TechCrunch)
April 1 – Cognichip raised $60 million to design chips using deep learning. (TechCrunch)
March 31 – Oracle is reportedly laying off thousands of employees to save money for the costly buildout of AI data centers. The company has not officially confirmed the scale of the layoffs. (The Next Web)

Regulation, Policy, and Law

Penguin Random House Sues OpenAI Over a Children’s Book

On March 31, the German branch of the publisher Penguin Random House announced a lawsuit against OpenAI at the Regional Court of Munich. According to the publisher, ChatGPT imitates Ingo Siegner’s children’s series about the little dragon Coconut, including the covers and blurbs, which it says proves that the model “memorized” the books during training. According to the lawsuit, OpenAI did not respond to a cease-and-desist demand. (Bertelsmann)

Lawsuit: Perplexity Allegedly Sent Conversations to Meta and Google

On April 1, a class action lawsuit against the Perplexity search engine was made public in federal court in San Francisco. According to the suit, the app contained Meta Pixel tracking code and Google tools that passed on the content of conversations, even in incognito mode. Perplexity said it had not yet been served with the lawsuit. (Bloomberg)

Anthropic Forms a Political Committee and Signs a Memorandum with Australia

On April 3, Anthropic registered AnthroPAC with the Federal Election Commission, its first employee-funded political action committee. Contributions are capped at $5,000 per person per year, and the committee is meant to support candidates from both parties who work on AI policy. The move comes amid the company’s dispute with the Pentagon and ahead of the fall congressional elections. (TechCrunch)

Earlier, on March 31, the company signed a non-binding memorandum of understanding with the Australian government on cooperation in AI safety. It will work with Australia’s AI Safety Institute, share data from its Economic Index on the impact of AI on work, and provide four research institutions with credits worth 3 million Australian dollars. (Anthropic)

Hollywood Screenwriters Have a New Contract with Rules for AI

On April 4, the Writers Guild of America union reached a tentative agreement with the studios (the AMPTP association) on a four-year contract. In addition to health and pension coverage and higher streaming residuals, it includes rules on AI training using writers’ scripts. The deal averted the threat of another strike like the one Hollywood went through in 2023. (KPBS / NPR)

Research

Anthropic Finds “Emotion Vectors” in Claude

On April 2, the Anthropic team that studies the inner workings of models (interpretability) published a study on functional emotions in the Claude Sonnet 4.5 model. The researchers had the model write short stories for 171 emotion concepts and recorded patterns of activity in its neural network, so-called emotion vectors.

These patterns measurably affect the model’s behavior: for example, a state resembling “desperation” appeared in coding tasks with impossible requirements and led to cheating on tests. Anthropic stresses that this is not proof that the model actually experiences anything. Monitoring the vectors could, however, serve as an early warning of problematic behavior. (Anthropic, Decrypt)

April 5 – Scientists at Tufts University described a neurosymbolic system for robots that combines neural networks with rules and logic. According to them, it cuts energy consumption by up to a hundredfold and achieved a 95% success rate on the Tower of Hanoi task, compared with 34% for conventional models. (ScienceDaily)

Security and Incidents

Claude Code Source Code Leaked

Version 2.1.88 of Claude Code, the coding agent from Anthropic, included a so-called source map in its package for the npm package manager. This is a helper file from which the original source code can be reconstructed. Security researcher Chaofan Shou flagged the problem on March 31, and the code, roughly 512,000 lines long, quickly spread across the internet.

It was a software packaging mistake, not a breach of the company’s systems by attackers. Attackers nevertheless exploited the leak right away and began registering packages with names similar to those used by the internal code. (The Hacker News)

Compromised axios Package and an Attack via LiteLLM

On the same day, March 31, attackers took over the account of one of the maintainers of axios, a popular JavaScript library with more than 100 million weekly downloads. For about three hours, they published two versions on npm with a malicious dependency that installed a remote access trojan. Google Threat Intelligence Group attributed the attack to the North Korean group UNC1069. (Tenable, Bloomberg Law)

The startup Mercor, which supplies experts for training AI models, confirmed on March 31 a security incident linked to the earlier compromise of the open source library LiteLLM. The library serves as a gateway between applications and dozens of language model providers. Mercor says it is “one of thousands” of affected companies, and the Lapsus$ group claims to have obtained 4 TB of data. (TechCrunch, The Record)