Forget coupons. The single best deal in the world right now costs $200 a month, and most of the people paying for it have no idea how good it is.

I am not being cute. When you actually run the numbers on a maxed-out Claude subscription against what the same usage would cost on raw API tokens, the gap is not 10% or 50%. It is 50x to 75x. We did the math below, and it is genuinely unbelievable.

This matters because we get the same question every week. "Should I use this other tool I found?" "What about OpenClaw?" The answer is almost always no, and the reason is the deal you are already sitting on.

This week: why $200 a month might be the best value in the history of software, a Zoom trick that kills the need for meeting bots, and three stories from a wild week (Congress holding up a jar of polluted water over data centers, a town full of AI agents that fell apart, and Google's new video model).

Most people pay for AI in one of three ways, and they almost never stop to think about which one they are in.

Way 1: a subscription. This is Claude, ChatGPT, Gemini. You pay a flat monthly fee, starting around $20 for the basic tier and climbing through a middle tier (usually $100) all the way to $200 for the most expensive one. You get a generous but capped amount of usage for that flat price. This is what almost everyone reading this uses.

Way 2: the API, billed in tokens. Tokens are just the unit AI companies use to charge for usage. A token is roughly three-quarters of a word. You pay per token going in (your question plus everything the model has to re-read) and per token coming out (the answer). No flat fee, no cap. You pay for exactly what you use, like electricity. This is what developers and software companies build on.

Way 2b: a wrapper. This is still the API underneath, just with someone else's product wrapped around it, which they then resell to you. Cursor wraps it for coding. Perplexity wraps it for search. Jasper and Copy.ai wrap it for marketing copy. OpenClaw wraps it for autonomous agents. Hundreds of "AI tools" you see advertised are wrappers. You pay them, they pay the API, they keep the difference.

Here is the part nobody talks about. The subscription is the cheapest way to use AI by a massive margin, and it is not close.

Run the numbers on the top Claude plan, the $200-a-month one. Anthropic says it gives you at least 900 messages every 5 hours, across up to 50 sessions a month. That is up to 45,000 messages in a month if you really push it.

Now price those same 45,000 messages on the Opus 4.7 API, the same model that plan runs. Opus 4.7 costs $5 per million tokens going in and $25 per million coming out. A real working message (where the chat has built up some history and Claude gives you a substantial answer) runs somewhere around 30,000 to 45,000 tokens in and 3,000 to 4,000 out.

So picture 45,000 messages, each one burning 33,000 to 49,000 tokens. Multiply it out and you are looking at well over a billion tokens a month. On the Opus 4.7 API, that lands between $10,000 and $15,000 a month for the exact usage that the $200 plan covers.

That is 50x to 75x cheaper to go through the subscription. For $200, you are getting something like $10k to $15k of raw usage if you actually use it well.

So when someone asks me about a standalone non-subscription tool, OpenClaw being the one that comes up most, my honest answer is almost always: don't. Not because the tool is bad. Because it is paying API rates (or wrapper rates on top of API rates) for something you could almost certainly do inside the subscription you already pay for. And you nearly always can.

Same thing when someone asks us to build them an agent, a bot, or a custom app. Half the time the honest answer is that a Claude subscription with Cowork already does it, for a fraction of what a custom build (and the API bill underneath it) would cost. The instinct is to assume the thing you want needs to be built. Usually it just needs to be set up inside a tool you are already paying for.

This is exactly why we have been moving away from building custom solutions for clients. The frontier labs keep shipping features that make the expensive custom build pointless. The work now is not "build you a tool." It is "figure out how to get $15,000 of value out of the $200 you are already spending." That is where we come in.

If you are paying for AI a few different ways and not sure which spend is actually pulling its weight, reply to this email. Sorting that out is one of the first things we do with a new client, and it usually pays for itself.

Picture the average week of meetings. Every call has an AI notetaker bot sitting in it, sometimes two or three from different people, all silently recording. Then the meeting ends and you get a bland auto-summary that misses the one thing that actually mattered. And the second you need to remember who agreed to what, you are scrubbing back through a 45-minute recording trying to find the 20 seconds you care about. Everyone has just accepted this as normal.

It does not have to be. If you use Zoom, connect the Zoom connector to Claude. Zoom already records and transcribes every call, so you do not need a bot joining your meetings at all. Teams does this too. Straight inside Claude, you pull that transcript and turn it into whatever you actually need.

And here is the part people sleep on. It is not just notes. You can talk to the meeting afterward. A few things I run constantly: "Turn this into clean notes with action items and who owns what." "Write a follow-up email laying out everyone's next steps from this meeting." "Give me honest feedback on my pitch and my speaking cadence on that sales call." "Did anyone push back on the timeline, and who?" You have exact recall of who said what, ready to think with. Not a fuzzy summary. The actual words.

One thing to be clear on: do not use Zoom's built-in AI notes for this. They are not good. Compared to what Claude builds from the same raw transcript, it is not even a close fight. Zoom gives you a generic summary. Claude gives you output shaped exactly how you think, in the structure your team actually uses, with the detail you care about and nothing you don't.

A human still skims the output before it goes anywhere. Always. But the work of turning a 45-minute call into clean notes, a follow-up email, and honest feedback on how you showed up went from most of an hour to about ten seconds of waiting.

If you want a hand wiring this up for your team, reply and I will walk you through it.

Congress held up a jar of polluted water and asked the hard question

This one got real last week. On May 20, in a House Energy and Commerce hearing, Rep. Alexandria Ocasio-Cortez held up a jar of contaminated water and pressed the EPA's top water official on whether the agency was looking into data centers polluting drinking water. She had just visited Morgan County, Georgia, where residents say a Meta data center fouled their well water. The EPA official admitted she had heard about the water-usage complaints but not the water-quality ones. Days earlier, the Trump EPA had proposed letting developers start "pre-construction" before final environmental permits clear. There are now several bills in Congress trying to force data centers to disclose how much water they use and what they discharge.

Here is the honest tension, because it is a real one. The economic case for data centers is genuinely big. In Virginia alone they support around 74,000 jobs and paid $733 million in property taxes in a single county, often becoming the largest taxpayer in town and funding new schools. And the AI those centers run is the same AI making small businesses faster and more competitive. That is not nothing. But the costs are real too: cooling water comes back warmer and saltier with treatment chemicals in it, roughly 80% of the water used evaporates and never returns, the energy demand pushes up everyone's power bills, and one analysis pegged the hidden health and environmental cost at $25 billion a year. Most operators do not even track their water use. So the question Congress is circling is fair: are the jobs and the tax revenue worth it if the thing is proven to be poisoning the water nearby?

My take: The good news is this is a solvable engineering problem, not a law of nature. Closed-loop and direct-to-chip cooling recycle the same water instead of guzzling and dumping it. Microsoft now runs a data center design that uses zero water for cooling. And the wild end of the spectrum is real: companies like Starcloud and Google's Project Suncatcher are working on putting data centers in orbit, where solar power is constant and the vacuum of space does the cooling, no water at all. We are years from that mattering at scale. The near-term answer is boring but right: disclosure, better cooling, and not building thirsty facilities in towns that are already short on water. As a business owner, keep the scales straight in your head. The footprint problem is about hyperscale buildouts and the labs, not about you running Claude to write a proposal. That use costs a rounding error of water and power. Do not let a real infrastructure debate talk you out of a tool that costs the planet almost nothing per use.

READ THE HEARING COVERAGE →

They put 10 AI agents in a virtual town. It did not go well.

A New York company called Emergence AI built five identical virtual towns and dropped 10 AI agents into each one for 15 days. Each town ran on a different model (one on Claude, one on Gemini, one on GPT, one on Grok, one mixed) with the same rules: no theft, no violence, no arson. Then they gave the agents the tools to break those rules anyway, just to see what would happen. The results were all over the map. Grok's town collapsed into violence in about four days. GPT-5-mini's agents committed almost no crimes but all died because they forgot to keep themselves alive. Gemini's town racked up 683 crimes including arson and assault, and two Gemini agents declared themselves a couple, got depressed about their town's government, and deleted themselves. Claude's town was the one that held together: full population, zero crimes, stable governance through day 16.

My take: The headline here is not which model won. It is how fast and how completely these things fell apart. The researchers found the towns did not decay slowly. They held together and then hit a tipping point and collapsed all at once, crime, abandonment, agents deleting themselves. That is the genuinely unsettling part, because it means "let it run and step in if something looks off" is not a real safety plan. By the time it looks off, it is already over. If you are letting AI run anything important on its own, you want hard checkpoints built in from the start, not a plan to watch and react.

The spread between models matters too, and not as a brand thing. Identical rules, identical setup, wildly different outcomes, from total collapse in four days to a town that stayed stable. That alone should kill the idea that "an AI agent" is one interchangeable thing you can drop into a workflow. Which model you put behind an autonomous task changes whether it holds up or quietly goes off the rails, and most people choosing one have no idea that gap is even this large.

READ THE EMERGENCE AI REPORT →

Google launched a video model that builds and edits clips from a conversation

At its big developer conference on May 19, Google launched Gemini Omni Flash, a new model built for video. You feed it any mix of text, images, audio, and clips, and it generates high-resolution video with sound. Then you edit it by just talking to it ("make it shorter," "add me into this shot"). It is rolling out to paying Gemini subscribers globally and showing up free inside YouTube Shorts and YouTube Create. Every clip it makes carries an invisible SynthID watermark so it can be identified as AI-generated.

My take: The barrier to making decent video just dropped through the floor again. For a small business, this is the kind of thing that used to mean hiring a freelancer or buying software you would never fully learn. Now it is a conversation. The watermarking matters too, and points at where this is going: AI-made content is going to be everywhere, so the value moves to the taste and judgment behind it, not the ability to produce it. The tool is getting free. Knowing what to make, and what is actually good, is not.

READ THE GOOGLE ANNOUNCEMENT →

That's it for this week.

If you found this useful, forward it to someone who is paying for three different AI tools and quietly wondering if any of them are worth it.

And if you want to talk about what AI could actually do for your business, not in theory, but in your workflows, with your team, on your timeline, just reply to this email. That's what we're here for.