7 May 2026: Trump’s AI Safety Reversal, Anthropic Boosts Claude Code, and US-China Talks (AM)

The Trump administration has quietly reversed course on AI safety testing, Anthropic wrapped a developer conference with major upgrades for Claude Code users, and Washington and Beijing are considering the first formal diplomatic channel on AI risk. Here is what matters on the morning of 7 May 2026.

The Trump administration has signed voluntary agreements with Google DeepMind, Microsoft, and xAI to run government safety checks on frontier AI models, reversing a policy position it held just weeks ago. The shift follows Anthropic’s announcement that it would not release its Claude Mythos model, judging its advanced autonomous cybersecurity capabilities too dangerous for public deployment. Until this week, Donald Trump had dismissed AI safety testing as Biden-era overregulation. He had even renamed the US AI Safety Institute to the Center for AI Standards and Innovation, pointedly dropping the word “safety.” Mythos appears to have changed the calculation.

National Economic Council Director Kevin Hassett has said Trump may issue an executive order mandating government testing of advanced AI systems before release. The voluntary agreements signed with Google, Microsoft, and xAI build on Biden-era frameworks that the Trump administration had previously sidelined. For UK observers, the shift matters. The UK government has been building its own AI safety framework through DSIT, and formal US testing requirements would bring Washington’s approach closer to European regulators, potentially easing compliance for AI companies operating across both markets.

Anthropic doubled the five-hour rate limits for Claude Code on Pro, Max, Team, and Enterprise plans and removed the peak-hours throttle that had frustrated developers during busy periods, announcing both changes at its Code with Claude developer conference. The conference also introduced a “dreaming” feature for Claude Managed Agents, currently in research preview. Dreaming is a scheduled process that reviews recent task sessions and identifies specific experiences worth keeping in long-term memory, helping agents running over hours or days carry context forward more reliably.

The compute news is substantial. Anthropic has signed an agreement to use the entirety of SpaceX’s Colossus 1 data centre, gaining more than 300 megawatts of capacity covering 220,000 NVIDIA GPUs. This joins existing deals with Amazon, Google and Broadcom, Microsoft and NVIDIA, and Fluidstack. For Claude Pro and Max subscribers, the practical result is improved availability and higher API rate limits for Opus models.

Washington and Beijing are weighing the creation of recurring talks on the security risks posed by AI models, with US Treasury Secretary Scott Bessent understood to be leading the American side. The Wall Street Journal reported the development this morning. No formal channel has been agreed and Beijing has not yet named a counterpart, but the proposal reflects growing anxiety in both capitals about frontier model capabilities, particularly in cybersecurity and autonomous systems.

For the UK, which maintains close ties with the US while aligning with EU AI governance, formal US-China dialogue on AI risk would be a significant development. The UK’s AI Safety Institute has already engaged with counterpart bodies in both countries, and any bilateral framework would provide clearer international context for UK domestic policy.

A TechCrunch analysis raises a pointed question about Elon Musk’s AI company xAI: is it really an AI lab, or is its true business selling data centre capacity to other companies? The piece examines xAI’s Colossus facility in Tennessee, its growing infrastructure footprint, and the economics of running large-scale GPU clusters. The argument is that renting spare capacity to other AI developers, in the manner of companies like CoreWeave and Lambda Labs, may generate more predictable revenue than model development alone. If the analysis holds, it has implications for how xAI is valued and whether developers might eventually access its infrastructure independently of the Grok product line.

Google has updated its AI Mode search experience to incorporate quotes and opinions from Reddit and other forums when answering questions that benefit from community knowledge. UK users searching for practical queries, such as which broadband provider is most reliable in their area or how to handle a specific tax form, may now see AI summaries that include real user experiences rather than only official sources. Google describes this as adding expert community perspectives to search. The update is live in AI Mode, accessible via Google Search for users in supported regions.

Here is everything else worth knowing from this morning’s AI news.

• DeepSeek could reach a $45B valuation: The Chinese AI lab is in its first external investment round, with sources telling TechCrunch the valuation could hit $45 billion. DeepSeek gained global attention in early 2025 after releasing a model that matched frontier performance at a fraction of the cost. [6 May]
• Apple to pay $250M to settle Siri lawsuit: Apple has agreed to pay $250 million to settle a class action lawsuit brought by US users who argued the company overpromised the timeline for Siri’s AI feature upgrades. [6 May]
• Genesis AI unveils GENE-26.5 robotics model: The Khosla-backed startup, which raised $105 million at seed stage to build foundational AI for robotics, has released its first model and a demo showing a set of full-stack robotic capabilities. [6 May]
• SpaceX mulling a $119B chip factory in Texas: A proposal submitted to Texas authorities describes a “multi-phase, vertically integrated semiconductor manufacturing and advanced computing fabrication facility.” If built, it would be one of the largest chip manufacturing investments in US history. [6 May]
• Snap and Perplexity’s $400M deal ends: The integration that would have put Perplexity’s AI search engine directly into Snapchat has been called off by mutual agreement, less than six months after the deal was announced. [6 May]
• Tinder owner slowing hiring to fund AI tools: Match Group, which owns Tinder, told investors it is slowing its hiring plans for the rest of the year because AI tools “cost a lot of money.” The company said AI is replacing some functions previously handled by new hires. [6 May]
• Scale AI wins $500M Pentagon contract: The Meta-backed data labelling and AI services company has secured a $500 million contract with the US Department of Defense to help analyse data and support decision-making. The deal is five times larger than Scale’s previous Pentagon contract. [6 May]
• Anthropic Managed Agents “dreaming” in preview: Alongside the Claude Code announcement, Anthropic confirmed that the dreaming feature for Managed Agents, which helps multi-step agents consolidate and retain relevant memories, is now live in research preview on the Claude Platform. [6 May]

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.