This Week in AI: May 25-31, 2026 (Week 22)

News of the week: May 25-31, 2026 (Week 22)

The week belonged to Anthropic, which released Claude Opus 4.8, promised to make its more powerful Mythos-class models available soon, and on the same day closed a funding round that made it the most valuable AI startup in the world. OpenAI published a framework that ties its safety practices to California law and the EU AI Act. Sam Altman admitted he was wrong in his predictions about AI’s impact on jobs, while Uber showed how quickly spending on AI tools can get out of hand. On the legal front, CNN sued the search engine Perplexity, and China restricted foreign travel for its leading AI experts.

New Models

Anthropic releases Claude Opus 4.8 and prepares Mythos for everyone

Anthropic launched Claude Opus 4.8 on May 28, succeeding version 4.7 from mid-April. API pricing stays the same as its predecessor: 5 dollars per million input and 25 dollars per million output tokens (a token is a small piece of text the model works with, roughly part of a word). Gizmodo

Anthropic itself describes the improvements as modest but noticeable. The biggest gain is in the model’s honesty: according to internal tests, the new Opus overlooks bugs in its own code about four times less often than Opus 4.7. The Next Web

The model comes with an accelerated Fast mode, which according to VentureBeat is significantly cheaper than the same mode in the previous version. VentureBeat
A feature called dynamic workflows is in preview, letting up to hundreds of helper agents (sub-workers that split a task between them) run in parallel on complex programming jobs. Gizmodo Japan
According to Reuters, the company also announced that in the coming weeks it will make its more capable Mythos-class models available to all customers. Because of cybersecurity risks, Claude Mythos Preview is so far reserved for a small group of organizations in Project Glasswing. KSL (Reuters)

Products and Agents

Microsoft Copilot Studio: computer-using agents are ready

On May 26, Microsoft published the May update of Copilot Studio, the platform where companies build their own AI assistants. So-called computer-using agents, which can click and type in ordinary apps and websites just like a person, have moved out of preview into general availability. Microsoft

The update also brought a new experience for workflows and voice interaction with agents in real time. Microsoft
For developers: on May 26, Google switched the default response format of the Interactions API for Gemini models to a new schema. This is a breaking change, and the old format is set to disappear on June 8. Google AI for Developers

Business and Investment

Anthropic is valued at 965 billion dollars and overtakes OpenAI

On May 28, Anthropic announced a 65 billion dollar Series H funding round that values the company at 965 billion dollars. That puts it ahead of OpenAI, whose March round valued it at 852 billion dollars. Bloomberg Law

The round was led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital. AP via Click2Houston
The total includes 15 billion dollars previously committed by cloud providers, 5 billion of it from Amazon. Built In
The previous Series G round valued the company at “only” 380 billion dollars, so its value more than doubled in just a few months. Techzine

Uber burned through its annual AI tools budget in four months

On May 26, Uber COO Andrew Macdonald admitted that spending on tokens is getting harder and harder to justify. According to reports, the company had used up its entire budget for this year on AI coding tools such as Claude Code and Cursor by April. Fortune

According to Macdonald, there is still no clear link between token consumption and value for customers. The case has become a symbol of the debate over how much companies actually pay for AI and what they get in return. Quartz

Earnings: Salesforce and Marvell grow thanks to AI, but their shares fall anyway

On May 27, Salesforce reported quarterly revenue of around 11.1 billion dollars and profit above analyst expectations. Annual recurring revenue from its AI agent platform Agentforce grew to roughly 1.2 billion dollars, but the stock weakened on a cautious outlook. SiliconANGLE
Chipmaker Marvell reported record revenue of 2.42 billion dollars on May 27, up 28% year over year. Growth was driven by custom chips and networking for data centers, which brought in 1.83 billion dollars. Invezz
Marvell expects revenue from custom chips to more than double next fiscal year. Even so, its shares and those of chip design software maker Synopsys closed the day lower. SiliconANGLE

Altman: I was pretty wrong about the impact on jobs

Speaking on May 26 at a conference hosted by Australia’s Commonwealth Bank, OpenAI CEO Sam Altman said he does not believe AI will cause a “jobs apocalypse”. The company was “roughly right” about the technology, he said, but “pretty wrong” about its social and economic effects. Euronews

Altman said he was “glad to be wrong” and that AI has so far not wiped out as many office jobs as he had feared. Fortune noted that this is a reversal from his warnings last year, and that Anthropic CEO Dario Amodei is softening his tone in a similar way. Fortune

Regulation and Law

CNN sues Perplexity

On May 28, CNN filed a lawsuit in federal court in New York against the AI search engine Perplexity. It claims the company copied its articles, videos and photos for its AI products and distributed content “identical or substantially similar” to the originals. Al Jazeera

Besides copyright infringement, the lawsuit also alleges misuse of the CNN trademark. A Perplexity spokesperson responded that “facts cannot be copyrighted”. NPR
CNN thus joins The New York Times, the publisher of The Wall Street Journal, the New York Post and Encyclopaedia Britannica, all of which have sued Perplexity before. NPR

OpenAI publishes its Frontier Governance Framework

On May 28, OpenAI published its Frontier Governance Framework. The document explains how its safety practices line up with new legal requirements, specifically California’s Transparency in Frontier AI Act and the Code of Practice for the EU AI Act. OpenAI

The document describes risk assessments in areas such as cybersecurity, chemical and biological weapons, manipulation and loss of control over a model. It also covers incident reporting and the involvement of outside experts. OpenAI

China restricts foreign travel for AI experts

According to a Bloomberg report from May 26, China has extended its foreign travel restrictions to leading AI experts at private companies, including Alibaba and DeepSeek. Selected employees now need approval to travel abroad, as the state seeks to protect technologies it considers strategic. Bloomberg Law

According to reports, authorities are also vetting potential investors and warning headhunters against poaching talent from these companies. ThePrint

Security and Incidents

AI agent carried out a cyberattack in under an hour

On May 28, Security Magazine described a case detected on May 10 by the research team at Sysdig: a cyberattack run by an autonomously acting AI agent. The whole operation took less than an hour. Security Magazine

The case shows why companies like Anthropic and OpenAI place so much emphasis on the cyber risks of their latest models. Attackers using AI can complete in minutes tasks that used to take a team of people days. Security Magazine