AI News08 Jul 2026
1. Anthropic built a tool that reads Claude's unspoken thoughts
Anthropic just showed off J-Lens, a tool that catches what Claude is thinking before it says a word. Ask the model to silently pick a sport and J-Lens reads "soccer" out of a hidden workspace that makes up less than a tenth of its activity; swap "spider" for "ant" inside it and the answer flips from eight legs to six. Researcher Neel Nanda calls the work fantastic but warns it throws false positives, so don't treat it as gospel. Still, if we can watch an AI think something it never says out loud, we might finally catch one hiding something.
AI News08 Jul 2026
2. OpenAI's most powerful model needed a government green light
OpenAI kept its strongest model yet locked behind a government review since June 26, open to just 20 vetted partners. Now the US Commerce Department has cleared it, and Sol goes public July 9, running up to 750 tokens per second. It's sharp enough at finding software vulnerabilities that officials wanted to vet it first. Sit with that for a second: a private company's product had to clear the state before you could touch it. Smart safety call, or a precedent you should worry about?
AI News08 Jul 2026
3. Ukraine will only run AI it can switch off itself
Ukraine just made a call the rest of us should study: it will only use AI it can run on its own servers, and that rule now covers all three fronts it leans on - government services, business, and the military. The trigger was the US ordering Anthropic to cut access to its best models. "AI sovereignty isn't just a defensive talking point, it's a necessity," says Roman Kyslyi, Ukraine's chief AI officer. Self-hosted models rarely match the frontier ones in the cloud, but when someone else can pull your plug mid-war, those last few points of performance stop mattering. Worth asking about your own stack too.
AI News08 Jul 2026
4. Claude Cowork now keeps working after you close your laptop
Claude Cowork jumped from desktop-only to web and mobile, and here's the part I like: you can kick off a task, shut your laptop, and the agent keeps going in the cloud. Set Monday's client prep for 6 am and it works through the email threads and transcripts and leaves the briefing drafted before you wake up. Anthropic looked at 1.2 million sessions from over 600,000 organizations and found a third of them are running business operations, not code. This is the coding agent quietly moving into the rest of the office, and it's on the Max plan first.
AI News08 Jul 2026
5. Chinese AI is undercutting OpenAI and Anthropic by up to 9x
Here's the number that should make OpenAI and Anthropic nervous: the same workload that costs $4,811 on Claude runs for $544 on China's GLM model, nearly 9x cheaper. Open Chinese models now run 60 to 90 percent below the US frontier, and GLM 5.2 landed within a single point of Claude Opus 4.8 on a key agentic benchmark at a fifth of the cost. In its first week, Z.ai's daily token volume grew about 27x and its customer count about 80x. US companies are doing the math, and the price gap is starting to pull real workloads east faster than we expected.