An AI model that beats or ties human experts 82% of the time just launched, yet no one’s paying attention to that. 

AI just embedded itself directly into Excel. 

And the CEO of the world's most valuable company called an open source project the single most important software release ever.

That's not even the biggest story of the past week. 

The biggest? Anthropic is suing the US government after the Pentagon officially labeled the AI lab as a supply chain risk. 

If your AI strategy doesn't account for what happened this week, you're already behind the people who do.

Here's what you need to know.

The Department of Defense officially designated Anthropic and its AI products a national security supply chain risk this past week, immediately banning federal use of Claude models.

This marks the first time the designation has ever been used against a domestic US company. Federal departments have six months to transition off Claude.

Anthropic CEO Dario Amodei responded by announcing the company will sue the US government, arguing the designation was meant for foreign adversaries, not American firms. Anthropic followed up on that promise Monday by officially filing suit.

The conflict started when Anthropic refused Pentagon demands to remove safety guardrails blocking Claude from fully autonomous weapons and mass domestic surveillance. The Pentagon wanted unrestricted access for all lawful use. Anthropic wouldn't budge.

Then a leaked internal memo from Amodei poured fuel on the fire. He criticized OpenAI for quickly agreeing to Pentagon terms, calling their approach safety theater. Amodei later apologized for the tone but reaffirmed Anthropic's safety commitments.

OpenAI, meanwhile, signed its own Pentagon deal and moved to fill the gap left by Anthropic's ban.

What it means: This feels like a broader application of the supply chain risk designation than what it was originally designed for.

But it's up to the current administration to decide how they want to apply certain laws. And that's exactly why we have a judicial system to begin with.

The Pentagon's Anthropic ban sparked concern across enterprises. But the three biggest cloud providers say most users won't feel a thing.

Microsoft confirmed it will continue offering Claude models through M365, GitHub, and AI Foundry for enterprises and general business users. Google announced Claude stays accessible on Google Cloud for non-defense workloads with no changes expected for other clients.

AWS told CNBC that its customers and partners can keep using Claude for non-defense projects without interruption.

So if you work in government but aren't on a Department of Defense contract, access stays intact. At least for now.

All three companies are investors in Anthropic and provide Claude models to their clients, which include major government agencies. They're framing the restriction as narrow and specific to defense-related contracts and direct DoD contractors only.

What it means: This will probably shake out in the courts and take a while. The hunch is it won't impact a whole lot of people unless you actually work in the Department of Defense.

But we're not lawyers. We just talk AI every day.

OpenAI dropped GPT-5.4 Thinking and GPT-5.4 Pro this past week, making a serious play against Anthropic and Google.

GPT 5.4 Pro now achieves an 82% win or tie rate against human experts on the GDPval benchmark. That means it matches or beats industry professionals at real-world tasks 82% of the time.

On most scientific benchmarks, GPT 5.4 is outperforming Claude Opus 4.6 and Gemini 3.1 Pro. Even in categories like computer use, tool calling, and real-world tasks where those competitors typically dominate.

It also features a one million token context window in the API, a 33% reduction in hallucinations, and native computer use capabilities.

The rollout was a little confusing though. OpenAI released GPT 5.3 Instant for free users one day. Then jumped straight to 5.4 the next. No 5.3 Thinking. No 5.3 Pro. Straight to 5.4.

What it means: Every few weeks, the crown changes hands. That's exactly why every business leader needs to stick with their AI operating system of choice.

You can't switch workloads every single week just because there's a new world's best model. Because there will be. Every. Single. Time.

Yuuuuup. ChatGPT is literally inside Excel now.

OpenAI announced ChatGPT for Excel this past week, rolling it out in beta. The new add-in embeds ChatGPT directly into Excel workbooks, letting users build, update, and analyze models using plain language instead of manual formulas.

Powered by GPT 5.4, the tool nearly doubles performance on internal investment banking benchmarks compared to previous versions. Users can also pull in trusted financial data from providers like FactSet, Dow Jones, LSEG, S&P Global, and Moody's without ever leaving the spreadsheet.

The tool can explain its outputs, trace errors, and link answers to specific cells. So teams can actually audit and verify results before making decisions.

Access is currently in beta for ChatGPT Business, Enterprise, EDU, Teachers, Pro, and Plus users in the US, Canada, and Australia. A Google Sheets version is coming soon too.

What it means: Microsoft rolled out Copilot in Excel a couple months ago. Claude's plug-in started going broad in February. And now OpenAI is following right in Anthropic's footsteps.

The spreadsheet AI race is officially three-deep. The question is which one to bet on.xxxx

Google launched Gemini 3.1 Flash-Lite this past week, the fastest and most cost-effective model in the Gemini 3 series.

How fast? Try 2.5 times faster time to first token than Gemini 2.5 Flash, pushing 363 tokens per second compared to 249. This model was purpose-built for real-time applications like customer support, content moderation, and user interface generation where even a two-second delay kills the experience.

Flash-Lite also introduces adjustable thinking levels, letting developers balance speed and reasoning depth depending on the task. Rapid classification on one end. Complex code generation on the other.

On benchmarks, it earned an ELO score of 1,432 on Arena with standout results in scientific knowledge, multimodal tasks, and multimodal Q&A.

The pricing is wild. Just 25 cents per million input tokens and $1.50 per million output tokens. That's roughly one eighth the cost of Gemini 3.1 Pro.

What it means: You probably shouldn't be using Opus 4.6 or GPT 5.4 Thinking for things like customer service. For certain high-volume tasks, Flash-Lite is a must-have.

If you built something six months ago, this model has to be on the radar at one eighth the cost of Pro.

NVIDIA CEO Jensen Huang made a bold declaration at the Morgan Stanley Technology, Media and Telecom Conference this past week.

He called OpenClaw the single most important release of software probably ever.

That's the CEO of the most valuable company on the planet praising an open source project that barely anyone knew about three months ago.

OpenClaw achieved in just three weeks the kind of adoption that took Linux 30 years. The project went through names like Clawdbot first, then Moltbot after getting a legal letter from Anthropic, before landing on OpenClaw. OpenAI recently acqui-hired the project's sole developer.

So why does the GPU king care about an autonomous agent platform? These agents consume 1,000 to one million times more tokens than standard chat interactions. That's unprecedented demand for compute infrastructure. Which means more demand for NVIDIA's GPUs.

What it means: Huang's comments signal that AI is shifting from information retrieval to autonomous task execution. And it's already reshaping how businesses use AI.

Even the big companies are scrambling to make their products more autonomous to compete with OpenClaw's approach. That tells the whole story.

According to The Information, OpenAI engineers have faced multiple GitHub outages lasting several hours, disrupting their ability to commit and collaborate on code. So the company is reportedly building its own code repository platform.

The project is still early and expected to take several months. OpenAI is weighing whether to offer it to customers or keep it exclusively for internal use.

This gets interesting when you look at the bigger picture though y'all. GitHub is owned by Microsoft, a major OpenAI partner. But that relationship has gotten more complicated. Since OpenAI's transition to a public benefit corporation, Microsoft has been investing in Anthropic and incorporating more Claude models into its own products.

OpenAI has similarly partnered with Microsoft competitors. So building a competing code platform fits the pattern of both companies slowly expanding into each other's territory.

What it means: With the recent explosion of autonomous coding tools like Claude Code and OpenAI's Codex, GitHub's servers are probably melting under the pressure.

Agents push commits around the clock now, even when nobody is sitting in front of them telling them to. You can see why both sides feel the squeeze.

Anthropic released a study this past week that warns AI is on track to disrupt a wide range of white-collar jobs.

AI models like Claude are already capable of performing up to 94% of tasks in computer and math roles. But they're currently only being used for about a third of those tasks in real-world settings.

That gap between what AI can do and what it's actually doing? The researchers describe it as vast.

The workers most at risk aren't in manual labor. They're highly educated, well-paid professionals. Lawyers, financial analysts, software developers. The study found those in the most AI-exposed jobs are 16 percentage points more likely to be female, earn 47% more on average, and are nearly four times as likely to hold a graduate degree.

For jobs requiring a physical presence, like cooks, mechanics, and bartenders, AI exposure remains near zero.

The study warns of a possible great recession for white-collar workers if AI adoption accelerates. Recent US labor data already shows a 14% drop in job finding rates for young workers since the rise of tools like ChatGPT.

What it means: This gap between AI capability and actual usage is something we've been talking about for a very long time. The exposure to high-income professional services was even predicted on this show two years ago.

There's a deeper dive coming on tomorrow's show because the findings go way beyond what a short recap can cover.