The Week Claude Was Illegal to Use.

Here is a sentence that should not be possible to write about software, and I am writing it anyway. For 18 days, the most advanced version of a major AI model was against the law to use inside the country that built it. Not because it did anything wrong. Because the US Commerce Department decided it was too powerful to let out the door without a permission slip, the same category of control usually reserved for jet engine parts and encryption chips headed overseas. This week, that ban got lifted. I want to walk you through what actually happened, because if you are the kind of person who still feels a step behind on all of this, this is the story that tells you the ground is moving faster than the headlines make it sound, and also that you are not the one who is behind. The people running this stuff do not fully agree on the rules either.

What actually happened with the Claude export ban?

Anthropic's Claude Fable 5 and Claude Mythos 5, its two most capable models, got hit with US export controls by the Commerce Department. For 18 days, using the top-tier version of Claude inside the United States was treated the same way as shipping a controlled piece of hardware overseas without a license. This week, those controls were lifted and the models became legal to use again, no restrictions.

I want you to sit with how strange that actually is, because the headline moves fast and the strangeness gets lost. We are used to the government regulating what leaves the country, not what a person inside the country is allowed to run on their own laptop. For those 18 days, the most capable AI model from one of the leading American labs got treated less like software and more like a jet engine part. That tells you the people writing the rules are genuinely unsure where AI capability belongs in the law, and they are figuring it out in public, in real time, with real consequences for anyone who had already built a workflow on top of it.

If you run a business and you were leaning on Claude for anything serious during that window, older, less powerful versions kept working the whole time. It was a narrow ban on the frontier tier, not a blackout of the whole product. But narrow does not mean small. It means the newest, best version, the one companies were racing to adopt, was the one pulled off the table.

Why did OpenAI offer the government a $42.6 billion stake, and why did Anthropic say no?

OpenAI pitched the US government a 5 percent equity stake in the company, worth roughly $42.6 billion, structured through a public wealth fund. Anthropic turned that idea down and proposed something different: a "digital dividend," a tax-and-payout mechanism that sends money to people directly instead of parking an ownership stake inside a government fund.

Strip the jargon out and you have two very different bets on how this era gets paid for. OpenAI's version hands Washington a seat at the table, an institutional stake in the upside. Anthropic's version tries to route the upside straight to households. Neither one is law. Both are opening offers in a negotiation that is going to shape who benefits when AI actually starts generating the kind of wealth these companies keep promising. I am not going to pretend I know which one wins. I am going to tell you to watch this fight closely, because whichever structure sticks becomes the template for the next decade of how AI money moves, and that affects you whether or not you ever type a prompt into anything.

This ties directly to the phrase I keep coming back to on this whole site: AI for everyone, not just the wealthy. A $42.6 billion equity stake sitting inside a government fund is still, structurally, wealth concentrating at the top, just a different top. A digital dividend paid to people is at least aimed in the other direction. Watch what actually gets built, not what gets pitched.

For 18 days, the newest version of a leading AI model was illegal to use in the country that built it. That should tell you something about how unsettled the rules still are, at every level, including the top.

Is AI agent progress actually slowing down?

Not in the way it sounds. Mark Zuckerberg told Meta staff at an internal town hall that AI agent progress "hasn't accelerated" the way the company expected. That is a real admission from a company spending enormous money on this bet, but it is a scaling and management problem inside one very large organization, not evidence that the underlying technology has stalled.

Here is the part that matters if you run a small business or work solo: a company with 70,000 employees has to move an agent idea through committees, legal review, brand approval, and a dozen VPs before it ships anything. A solo operator, or a five-person shop, has none of that. You can deploy one working agent, an AI that answers your phone, follows up your leads, or drafts your invoices, in an afternoon, while a company the size of Meta is still scheduling the meeting about whether to schedule the meeting. Their slowdown is a bureaucracy problem. Your speed is the actual advantage, and it is the whole reason I built my business the way I did.

Why is OpenAI shutting down Sora, and what does that tell you about the AI race?

OpenAI is shutting down Sora, its video-generation product, and reportedly walking away from a $1 billion licensing deal with Disney. Reporting frames the pivot as OpenAI redirecting toward enterprise products and a new custom chip, with Claude Code reportedly eating into OpenAI's standing among developers.

If a company can spend real money building a headline product, sign the kind of company as recognizable as Disney, and then walk away from both inside the same year, that tells you something honest: even the labs at the very top of this race do not know for certain which bets will land. Nobody has this fully figured out, not the billion-dollar companies, not the government, not you. That should take some of the pressure off. You are not supposed to have a perfect AI strategy in July 2026. Neither does OpenAI.

What just got a lot cheaper, and what does that mean for a small business?

Claude Sonnet 5 became the free default worldwide this week, with paid usage priced at $2 per million tokens in and $10 per million tokens out through August 31, 2026. Separately, xAI opened its Grok Voice Agent Builder to everyone, no code required, at 5 cents a minute with sub-second response latency.

Translate that out of tech-speak. A capable AI model, one that can write, summarize, and reason through a real business problem, now has a cost floor close to zero for casual use and a genuinely small bill for heavy use. A voice agent that can answer your phone and hold a real conversation costs about a nickel a minute to run, and you do not need to write a line of code to build it. The tooling barrier that used to separate "big company with an engineering team" from "regular business owner" is close to gone. What separates them now is not access to the tools. It is knowing which vertical, which script, and which follow-up sequence actually converts a lead into a customer. That is not a coding problem. That is a business problem, and it is exactly the gap I spend my time closing for people.

Does the "AI compute shortage" story still hold up?

Not as cleanly as it did a few months ago. Meta unveiled a new AI cloud-reselling business called Meta Compute this week. Meta's stock rose roughly 9 percent the same day two infrastructure-focused companies, CoreWeave and Nebius, fell 14 percent and 17 percent respectively.

For over a year, part of the pitch from AI infrastructure startups has been scarcity: not enough chips, not enough data centers, get in line and pay up. When one of the biggest tech companies on earth can casually stand up its own compute-reselling arm and the market rewards it while punishing the specialist infrastructure players, that scarcity story gets a lot harder to sell with a straight face. I am not telling you compute is suddenly unlimited. I am telling you to be skeptical the next time someone uses "shortage" as the reason their AI product costs what it costs.

Is it risky to use a cheap AI model just because it is cheap?

Yes, and this is the part I actually want small business owners to sit up for. Palantir CEO Alex Karp said this week that AI labs are chasing raw power over enterprise usefulness, while US businesses increasingly adopt cheaper Chinese AI models despite government restrictions against doing exactly that.

If you are a small business owner tempted by an AI tool that seems too cheap to be true, that is worth a second look before you plug your customer data into it. This is not me fearmongering about foreign technology. It is a straightforward point: restricted models carry restrictions for a reason, and a small business does not have the legal team to absorb a compliance mess if it goes sideways. Cheap is not the same as smart. I would rather point a client toward a model with a clean compliance story at $2 per million tokens than toward something unrestricted-sounding that quietly puts their business at risk.

Why does it matter that Google is losing talent to Anthropic?

Google is reportedly experiencing a talent exodus, with researchers including Noam Shazeer and John Jumper said to be headed to Anthropic, while Google's next model, Gemini 3, is reportedly slipping in coding benchmarks.

Here is a filter I would give anyone picking a long-term AI vendor for their business: watch where the talent moves, not just where the benchmark chart points this month. Benchmarks get gamed, marketed, and cherry-picked. People vote with their careers. When top researchers leave one lab for another, that is a slower, quieter, and honestly more reliable signal about who is actually winning the next round than any single leaderboard screenshot.

What did Claude Mythos find when it went looking for security holes?

Claude Mythos scanned open source code and confirmed 23,019 real vulnerabilities across 1,000 repositories, with a 90.6 percent independent confirmation rate. That is a large chunk of the free, open software quietly running behind ordinary websites, apps, and business tools, holding holes nobody had found until an AI went looking.

I want to be straight with you about what that number means both ways. It is genuinely useful: defenders now have a tool that can comb through code at a scale no human team could match, and 90.6 percent confirmed means it is not just flagging noise. It is also a reminder that the same capability cuts both directions. A tool that can find 23,019 real holes for the people trying to fix them can, in someone else's hands, find them for people trying to exploit them. That is not a reason to panic. It is a reason to take security seriously if any part of your business touches open source software, which, if you are honest about it, it almost certainly does.

So what do you actually do with all of this?

You do not need to track export control law, equity fights in Washington, or Google's org chart to run a good business. What you need is the plain-English translation, which is what this post is. Here is the short version. The rules are being written in real time, even by the people who built this technology, so do not feel behind for not having it figured out, because nobody does. Costs are dropping fast, Claude Sonnet 5 free worldwide, voice agents at 5 cents a minute, which means the tools are no longer the barrier. The barrier now is knowing what to build and where it converts. And the risk is not "AI is scary." The risk is plugging in the wrong tool, cheap and unrestricted, without understanding what you are exposing.

I have been coding since 1983, a kid on a Timex Sinclair 1000 with 2 kilobytes of memory, and I have watched enough technology cycles to know the pattern. The first wave scares people. The second wave prices out the people who waited too long to learn it. I built this whole site, and my Choose Your Hard framework, so that Gen X, Boomers, and anyone who feels like the twenty-somethings got a head start does not have to figure this out alone or from a headline written for engineers.

If you want the fuller three-lane recap of everything from today's show, including the real estate market read and the health and fasting lane, I put the full picture on the hub post over at connorwithhonor.com/blog. This post is the AI lane only, on purpose, because that is what this site is for.