On June 26, OpenAI announced GPT-5.6 in three versions: Sol, the new flagship; Terra, a balanced model for everyday work; and Luna, a fast, affordable one. At the US government's request, access starts with a limited preview for roughly 20 trusted partners, and the government will approve wider release customer by customer. It follows Anthropic switching off access to its newest models, Fable and Mythos, earlier this month under a government export directive (an order limiting where a technology can go).
Why it matters
The US government now sits directly inside the release process for the most capable AI models, a level of involvement that did not exist even a year ago. That affects when, and whether, the newest tools reach you and the companies you may work for.
In a letter to US senators made public this week, Anthropic alleged that operators tied to Alibaba's Qwen AI lab ran 28.8 million conversations through Claude using nearly 25,000 fraudulent accounts between April 22 and June 5. Anthropic calls it distillation, meaning training your own model on another model's answers, done here at industrial scale and aimed at Claude's coding, multi-step reasoning and cybersecurity skills. It says the campaign was larger than the three Chinese lab campaigns it flagged in February combined.
Why it matters
Distillation turns a rival's chatbot into a free teacher. Policing it is becoming a front line in the US-China AI race, and a test of whether AI companies can protect what they build.
Jonas Adler and Alexander Pritzel, both key contributors to Google's Gemini models, are leaving for Anthropic, TechCrunch reported on June 24. Their exits extend last week's wave, when transformer co-inventor Noam Shazeer left for OpenAI and Nobel laureate John Jumper joined Anthropic. With OpenAI and Anthropic preparing to go public, both can offer pre-IPO equity (shares granted before a company lists on the stock market) that Google cannot match.
Why it matters
One departure is noise, but a steady outflow of Gemini's core builders is a signal about where top researchers think the frontier is heading.
On June 22, a fiber cut in eastern North America triggered routing failures, breakdowns in how internet traffic finds its way, at Zayo, an internet backbone provider most people have never heard of. X, Reddit, Discord, Zoom, Canva and several Microsoft 365 services were unreachable for parts of the afternoon before recovering. The platforms' own servers were fine; the paths to reach them were not.
Why it matters
The internet's plumbing is more concentrated than it looks, and the apps you use every day depend on infrastructure companies you never see.
The National Highway Traffic Safety Administration, the US road safety regulator, opened a special crash investigation after a Tesla Model 3 drove through a brick house in Katy, Texas, killing 76-year-old Martha Avila inside. The driver told authorities the car was in self-driving mode; Tesla says he overrode the system by flooring the accelerator, reaching 73 mph. The victim's family has sued Tesla.
Why it matters
When driver-assistance software is involved, who is responsible (the driver, the carmaker or the code) is being worked out right now in courts and by regulators.
Open-source AI lab Reflection will pay SpaceX $150 million a month, starting July 1, to use Nvidia GB300 chips at the Colossus 2 data center near Memphis, a deal worth up to $6.3 billion if it runs through 2029. It comes a week after SpaceX agreed to buy Cursor maker Anysphere, and adds to Colossus deals renting computing power to Anthropic, Google and Cursor. Reflection was founded by two former DeepMind researchers and raised $2 billion at an $8 billion valuation last October in a round led by Nvidia.
Why it matters
Computing power, not ideas, is the scarcest resource in AI right now: even well-funded labs sign billion-dollar rent checks for someone else's chips.
Berlin-based Stark, founded in 2024, builds loitering munitions (one-way attack drones), uncrewed vessels and battlefield software. Its round, announced June 23 with the NATO Innovation Fund also taking part, roughly tripled its valuation to more than 3 billion euros, and over 80 percent of the money is set aside for expanding manufacturing and research. Reporting the same week questioned how well its drones performed in tests.
Why it matters
European defense tech now attracts Silicon Valley-sized checks, a clear read on where venture money flows when geopolitics heats up. The test reports are a reminder that big funding is not proof a product works.
The through-line this week is control. Washington now decides who can use OpenAI's newest models, Anthropic is policing who learns from Claude's answers, and Google is watching Gemini's builders leave for the two labs headed to public markets. Capability is no longer the only contest: access to models, to talent and to the answers rivals train on is being rationed, fought over in public and priced.
Underneath it all, the physical layer kept asserting itself. One cut cable stalled some of the world's biggest platforms, a $6.3 billion deal came down to renting chips, and a fatal crash in Texas showed that when software steers a car, accountability is still unsettled. The headlines are about models; the lasting stories are computing power, infrastructure and rules.
What to watch
OpenAI says GPT-5.6's broader release comes in the coming weeks, with the government approving access customer by customer; watch whether it, and Anthropic's paused Fable and Mythos models, reach the public in July.
Test yourself
A few quick questions on this week. Just for you, nothing is recorded.
1. Why can only a small group of organizations use OpenAI's GPT-5.6 at launch?
2. In the Anthropic-Alibaba dispute, what does 'distillation' mean?
3. On June 22, why were X, Reddit and Zoom unreachable even though their own servers were working?
4. What is Reflection AI paying SpaceX $150 million a month for?
