In partnership with

Anthropic just launched its best AI model yet and spent as much time explaining what it refuses to do as what it can do. Google quietly shipped three smaller models while its big upgrade sits behind schedule. Grok found its way into Google's own spreadsheets, and a team of 15 AI models tried, and mostly failed, to fly a drone through an office. If today has a theme, it's this: AI companies are getting more comfortable admitting where their tools stop, and that admission is worth paying attention to.

Thinking about hiring globally? Start with an EOR.

The best person for your next role might not live near your office—or even in the same country.

More companies are realizing they don't need to open entities everywhere just to access global talent. Instead, they're using EOR to hire internationally faster, stay compliant, and avoid building local infrastructure before they're ready.

Oyster's EOR helps companies hire, pay, and support employees in 180+ countries while Oyster handles payroll, compliance, taxes, and local employment requirements.

Anthropic's Newest AI Model Was Built to Say No to Hackers

What happened: Anthropic released Claude Opus 5 on July 24, cutting the price in half compared to its previous top model (now $5 per million input tokens and $25 per million output) while beating it on coding and knowledge benchmarks, according to Axios and Anthropic's own announcement. The catch: Anthropic says the model deliberately does not push the frontier on offensive cybersecurity, and it triggers safety checks 85 percent less often than its predecessor by design, not by accident.

Why it matters to you: A cheaper, more capable AI model is good news for your budget. But this is also the clearest sign yet that AI companies are choosing, on purpose, how much capability to hand you, and holding back specific skills like hacking know-how rather than trying to make every model maximally powerful.

What to do about it: If you are already paying for a premium AI plan for basic tasks, check whether Opus 5's lower price covers what you actually use before renewing at the old rate.

Google Promised a Bigger AI Model. It Shipped Three Smaller Ones Instead

What happened: Google's next flagship Gemini model was reportedly running months behind schedule, specifically on coding, according to a TechCrunch report sourced to current and former employees. Days later, Google released three smaller models instead: Gemini 3.6 Flash, a cheaper everyday workhorse model, 3.5 Flash-Lite, built for speed, and 3.5 Flash Cyber, a security-focused model with restricted access, per Google's own announcement.

Why it matters to you: Google filled the gap with faster, cheaper tools while the model everyone actually wanted stays in testing. If you use Google's AI products for lightweight tasks, you get a price cut today. If you were waiting for a major capability jump, keep waiting.

Grok Just Moved Into Your Spreadsheets

What happened: SpaceXAI released a free Grok add-on for Google Workspace on July 24, embedding it directly into Google Sheets, Docs, and Slides. In Sheets, it reads selected ranges, cites the exact cells behind its answers, and writes formulas. In Slides, it builds a full deck from an outline, according to x.ai's own announcement.

Why it matters to you: If your business already runs on Google Workspace, you now have a capable AI tool sitting inside the software you already pay for, with no new subscription required.

What to do about it: Install the add-on and test it on one recurring spreadsheet or deck task this week before you pay for a separate AI tool to do the same job.

Fifteen AI Models Tried to Fly a Drone. Most Couldn't Find the Room

What happened: Anthropic's research team, working with Andon Labs, tested 15 AI models on a simple task: fly a real quad-rotor drone through an indoor office, locate a specific person, and follow them. Newer models cleared most of the individual steps, but according to Anthropic's own research writeup, even the strongest model failed the first step, mapping the room, and that early failure broke every step that depended on it.

Why it matters to you: AI keeps getting better at reasoning and planning on a screen, but basic real-world perception, understanding a physical space, is still a weak point even for the best labs. Be skeptical of any product promising your business can run "autonomously" in the physical world today.

One AI Model Now Makes Images, Video, Audio, and Robot Instructions

What happened: Black Forest Labs launched FLUX 3 on July 23, a single model trained jointly to generate images, 20-second video clips with synced audio, and now action predictions used to control physical robots. A robotics offshoot called FLUX-mimic is already running on Audi's production line, according to VentureBeat.

Why it matters to you: The line between "AI content tool" and "AI that controls machinery" is blurring faster than most people realize. If your business touches manufacturing or physical operations, this is worth watching even though today's rollout is limited and gated.

The Bottom Line

None of this week's news is really about which model is smartest anymore. It is about which companies are honest about the ceiling on their own tools, whether that ceiling is a safety choice, a delayed release, or a drone that cannot map a room. The AI industry wants you to believe every release is a leap forward. Most of them are a trade-off dressed up as one. Know which trade-off you are actually making before you hand over your budget or your workflow.

Enjoying the Ride?

If this issue was useful, forward it to one person who needs to see it. And if someone forwarded this to you, subscribe so you get the next one directly.

Talk tomorrow,
Mark Shilensky