Anthropic Releases the Claude 4.6 Family with Major Gains in Reasoning and Coding
On May 6, Anthropic released a new model family consisting of Claude Opus 4.6, Sonnet 4.6, and Haiku 4.5. The improvements center on reasoning and coding performance, and Opus 4.6 in particular is said to show significant gains on complex tasks.
To mark the release, CEO Dario Amodei spoke with The Verge about the background of the Claude 4.6 launch and the competitive landscape. He also explained why Anthropic takes a cautious approach to Claude's autonomy, making the company's stance clear: it has no intention of having "Claude run your life." Amodei went further, predicting that AI will bring an end to the "age of discovery" and that the way we access information could fundamentally change.
Meanwhile, according to The Decoder, Anthropic has signed a contract to spend roughly $200 billion on Google Cloud over the next five years — more than 40% of Google's entire cloud backlog. OpenAI and Anthropic alone account for about half of the $2 trillion in committed cloud revenue across Amazon, Microsoft, Google, and Oracle, underscoring the enormous scale of AI infrastructure investment.
Google and Meta Get Serious About the Personal AI Agent Race
According to The Decoder, Google and Meta are each testing personal AI agents internally. Google's agent, codenamed "Remy," and Meta's "Hatch" both aim to autonomously handle everyday tasks.
Google has reportedly wound down its browser agent project "Mariner" to concentrate resources on this new integrated assistant. The market is shifting from agents that operate in the browser toward assistants built directly into email, calendars, and shopping platforms, leaving Google and Meta to chase Anthropic and OpenAI, which are currently out in front.
OpenAI and NVIDIA Push Gigascale AI Training Infrastructure with the MRC Protocol
On May 6, NVIDIA announced that it has added Multipath Reliable Connection (MRC) support to its Spectrum-X Ethernet networking platform. MRC is a protocol that distributes traffic across multiple network paths over a single RDMA connection, and OpenAI, Microsoft, and Oracle have already adopted it in production.
Sachin Katti of OpenAI said that "MRC deployment on the Blackwell generation was extremely successful," crediting it with substantially avoiding network-related delays and disruptions in frontier training. Some of the world's largest AI factories — including Microsoft's Fairwater data center and Oracle Cloud Infrastructure's Abilene data center — have adopted MRC, and the protocol has been published as an open specification through the Open Compute Project.
NVIDIA also says that when paired with multi-plane configurations, MRC maintains low latency even at scales of hundreds of thousands of GPUs, making it an essential infrastructure technology for next-generation AI supercomputers.
Google DeepMind Invests in Eve Online Developer as AI Research Reaches Deeper into Games
According to Bloomberg, Google DeepMind has acquired a minority stake in CCP Games, the developer of Eve Online. The plan is to train DeepMind's AI technology within the world of Eve Online — a sign that AI research is pushing even deeper into gaming.
Most AI Trading Bots Post Losses in Wall Street Contest
According to Bloomberg, the majority of systems lost money in a trading contest featuring the world's leading AI models. The results highlighted a fundamental gap between large language models and how markets actually work: the models traded too frequently, and even when given identical instructions, they made wildly different decisions.
Apple Quietly Ends High-Memory Mac Studio Configurations
According to reports on Reddit's LocalLLaMA community, Apple has been quietly discontinuing high-memory configurations of the M3 Ultra Mac Studio. The 512GB configuration was removed in March, the 256GB configuration is no longer offered, and 96GB is now the only option available. The Mac mini is also capped at 48GB, and a shortage of high-memory chips has been pointed to as the cause. For users running large models locally, losing access to a relatively affordable large unified memory environment is a major blow.
TikTok Wins Approval for $25 Billion Data Infrastructure Investment in Thailand
Thailand's government has approved a 958 billion baht (approximately $29 billion) investment project, anchored by an 842 billion baht data infrastructure expansion plan from TikTok System (Thailand). The project includes additional servers in Bangkok and neighboring provinces, expanded data storage and processing facilities, and digital literacy and e-commerce training programs. It is part of Thailand's push to become a regional hub for AI and cloud services.
New York Stock Exchange Gears Up for a Data Center IPO Rush
According to Bloomberg, Wall Street banks are preparing to raise billions of dollars through listings of data center companies. Next week's IPO of a Blackstone-sponsored data center acquisition company will serve as the starting signal, with roughly $7 billion expected to be raised over the next 18 months, including from Singapore's DayOne Data Centers. CSquare, backed by Brookfield, has also confidentially filed, and numerous other companies are said to be considering U.S. IPOs.