The Week in AI: March 23-29, 2026 (Week 13)

This week’s news: March 23-29, 2026 (Week 13)

The thirteenth week of the year showed how quickly AI is turning from a chat window into an agent that operates a computer on its own: Anthropic gave Claude control of the mouse and keyboard on the Mac and added an automatic approval mode for programmers. OpenAI, by contrast, made cuts and shut down its Sora video app, which also killed a billion-dollar deal with Disney. Google added the Gemini 3.1 Flash Live voice model and the Lyria 3 Pro music model, and Arm unveiled its own chip for the first time in its history. On the legal front, the week brought a landmark court ruling in Anthropic’s dispute with the Pentagon, a leak revealing the secret Claude Mythos model, and a European Parliament vote on delaying part of the AI Act.

New Models

Google: Gemini 3.1 Flash Live and Lyria 3 Pro

On March 26, Google introduced the voice model Gemini 3.1 Flash Live, which it calls its best audio model yet. It powers the conversational mode Gemini Live and the Search Live feature, a voice search with camera support that has also expanded to more than 200 countries (Google Blog).

Developers can access the model in preview through the Gemini Live API in Google AI Studio. All generated audio carries an invisible SynthID watermark, which makes it possible to tell later that a recording was machine-made (Gigazine).

A day earlier, on March 25, Google launched the music model Lyria 3 Pro. While the previous Lyria 3 could only produce thirty-second clips, the new version composes tracks up to three minutes long, and users can define the song structure, such as verses, choruses and a bridge (Google Blog).

Lyria 3 Pro is available to subscribers of the Gemini app, in the Google Vids tool and to businesses via Vertex AI; according to Google, it was trained on licensed data (Music Business Worldwide).

Open Voice Models from Mistral and Cohere

On March 26, France’s Mistral AI released Voxtral TTS, a text-to-speech model. It has roughly 4 billion parameters, supports nine languages and, according to the company, can clone a voice from a sample shorter than five seconds (SiliconANGLE).

The model weights, meaning the actual “learned” data, are freely downloadable, but only for non-commercial use under the CC BY-NC 4.0 license. Commercial customers pay for access via the API (TechCrunch).

On the same day, Canada’s Cohere introduced its first voice model, Cohere Transcribe, for speech-to-text transcription. It has 2 billion parameters, handles 14 languages and was released under the truly open Apache 2.0 license (The Decoder).

On the Open ASR Leaderboard it achieved an average word error rate of 5.42%, which according to Cohere puts it ahead of OpenAI’s Whisper Large v3 and ElevenLabs Scribe v2. However, it lacks timestamps and speaker diarization (Cohere Docs).

Products and Agents

Claude Takes Control of the Mac and Gets an Auto Mode

On March 23, Anthropic launched the computer use feature in Claude Cowork and Claude Code. On a Mac, Claude opens files, works in apps and the browser on its own, and users can even assign it tasks from their phone through the Dispatch feature (Claude Blog).

It is a research preview for Pro and Max subscribers. Claude first tries direct integrations with apps such as Slack or Google Calendar, and only when that is not possible does it “click” around the screen, which is slower (MacRumors).

A day later, on March 24, Anthropic added a so-called auto mode to Claude Code. Instead of asking the programmer about every file edit or command, a dedicated safety classifier assesses the risk of each action and approves lower-risk steps on its own (SiliconANGLE).

For now, the mode works with the Claude Sonnet 4.6 and Claude Opus 4.6 models, first for team accounts and gradually for enterprise customers and the API.

Gemini Courts Competitors’ Users

Around March 26, Google began offering the option in the Gemini app to import “memory” and conversation history from ChatGPT or Claude. Switching between assistants no longer means starting from scratch (PCWorld).

Agents for Businesses

March 23: Alibaba launched the Accio Work platform, where small and medium-sized businesses can build AI agents for trade in more than 100 countries without any coding (Reuters).
March 24: Oracle unveiled agentic features for its database, including the Autonomous AI Vector Database and the Private Agent Factory tool for building agents on top of company data (Oracle).
March 26: Workday announced hundreds of new agents for HR and finance and the general availability of its Sana assistant (Workday).

Business and Investment

OpenAI Shuts Down Sora, Disney Backs Out

The most discussed news of the week was OpenAI‘s announcement on March 24 that it would shut down the Sora video app and its API, only about half a year after its grand launch. The company said it wants to focus on tools for real-world, physical tasks such as robotics (Engadget).

Along with Sora, the December 2025 deal with Disney also fell through. Disney was supposed to invest one billion dollars in OpenAI and license more than 200 characters, but the investment never happened (PC Gamer).

According to an analysis by TechCrunch, the decision was driven mainly by the enormous computing costs compared to the app’s low revenue (TechCrunch).

OpenAI Foundation to Give Away a Billion Dollars

The company’s nonprofit arm, the OpenAI Foundation, announced on March 24 that it will distribute at least one billion dollars in grants over the next year. The money will go to four areas: life sciences and curing diseases, the impact of AI on jobs and the economy, societal resilience to AI risks, and community programs (Bloomberg).

The AI risk resilience area will be led by co-founder Wojciech Zaremba (Fortune).

Arm Makes Its Own Chip for the First Time

Britain’s Arm, which until now only licensed processor designs to other companies, unveiled on March 24 the first chip of its own in its 35-year history. The Arm AGI CPU server processor has up to 136 cores, is manufactured on a 3nm process at TSMC and targets data centers running AI agents (The Register).

The chip was developed in collaboration with Meta, which is its first customer. Arm shares rose after the announcement because the company promised billions of dollars in annual revenue from the new chip (The Star / Reuters).

Smaller Investments

March 25: London startup Granola, known for its app for AI meeting notes, raised 125 million dollars at a valuation of 1.5 billion dollars; the round was led by Index Ventures (Bloomberg).

Regulation and Law

Court Halts the Pentagon’s Penalty Against Anthropic

On March 26, federal judge Rita Lin in California issued a preliminary injunction suspending the designation of Anthropic as a “supply chain risk”. The Pentagon imposed this designation, until now reserved for hostile foreign companies, after Anthropic refused to allow Claude to be used for mass surveillance of Americans and for autonomous weapons (Axios).

The judge also blocked President Donald Trump‘s directive for federal agencies to stop using Claude. In her view, the government’s actions likely violate the law and amount to retaliation for the company making the dispute public, which the First Amendment to the Constitution prohibits (FedScoop).

The injunction is temporary and the judge delayed it from taking effect by a week, so the dispute is far from over.

European Parliament Wants to Delay Part of the AI Act

On March 26, members of the European Parliament approved, by 569 votes in favor, 45 against and 23 abstentions, their negotiating position on the Digital Omnibus package. It is intended to postpone obligations for so-called high-risk systems under the AI Act, such as AI used in hiring or biometrics (Biometric Update).

Parliament also proposed a new ban on AI systems that create sexual content without the consent of the people depicted. The vote opened negotiations with the Council of the EU, so this is not final approval.

Other Legal News

March 24: The city of Baltimore sued xAI, X and SpaceX because the Grok chatbot generated sexualized deepfakes of real people, including minors. It is the largest American city to sue over Grok (The Guardian).
March 24: Washington state Governor Bob Ferguson signed HB 2225, a law on so-called companion chatbots. They must remind users that they are not talking to a human and have procedures for detecting suicidal thoughts; the law takes effect on January 1, 2027 (Hunton).

Security and Incidents

Leak Reveals the Claude Mythos Model

On March 26, Fortune magazine reported that Anthropic had around 3,000 unpublished blog materials in a publicly accessible data store. Among them was a draft article about the as-yet unannounced Claude Mythos model, called Capybara in an earlier version (The Decoder).

The company attributed the leak to human error in the configuration of its content management system and blocked access. It confirmed the model’s existence and described it as a “step change” in capabilities; according to the leaked texts, it also brings unprecedented risks in the area of cybersecurity (Euronews).

Protecting AI Agents

March 23: Palo Alto Networks introduced Prisma AIRS 3.0, a security platform covering the entire lifecycle of AI agents. At its core is the AI Agent Gateway, which monitors what agents are allowed to access (PR Newswire).