Volume 42: Own the Routine, Rent the Frontier When It Counts

A public company just cut its AI bill nearly in half while usage kept climbing. The saving did not come from a bigger discount. It came from deciding which parts of its intelligence to own, and which to keep renting by the token.

🧭 Founder's Corner: Argues that every AI bill is a strategic choice whether you make it on purpose or not, and shows how to decide which parts of your intelligence are worth owning outright.

🧠 AI Education: Explains why the same technology built for coding is already handling ordinary business tasks, and the one setting that decides how much control you keep when you hand off work.

✅ 10-Minute Win: Walk away with a keep, cut, or consolidate verdict on every AI tool you pay for, plus a reusable audit you can run again next quarter.

Let's dive in.

1) OpenAI releases GPT-5.6 to the public after a government review

Summary: OpenAI launched its GPT-5.6 family publicly on July 9 in three tiers: Sol for the hardest work, Terra for everyday tasks at half the cost of the flagship, and Luna for speed and low price. The models sat behind a government-only preview for about two weeks until the Commerce Department finished additional testing and cleared a wider release. 

Why it matters: The tiered lineup means you no longer pay flagship prices for routine work, so match the model to the job. The bigger signal is the review itself. For the first time, a US lab held back a public launch pending a government check, which hints at how model releases may work from here.

2) Medicare proposes its first payment category for AI diagnostic software

Summary: In its proposed 2027 outpatient payment rule, CMS is creating a standardized payment pathway for algorithm-driven clinical software, renaming the category "Software as a Medical Service." It covers AI reading of retinal images, echocardiograms, bone scans, concussion assessments, and brain MRIs, with comments open until August 31. 

Why it matters: Reimbursement is the quiet gate that decides which AI tools actually reach patients. A tool can be FDA cleared and clinically strong and still go nowhere if nobody gets paid to run it, so this is the plumbing that turns approved AI into adopted AI.

3) Independent testers ranked Grok 4.5 fourth, and its hallucination rate doubled

Summary: xAI released Grok 4.5 on July 8, with Elon Musk calling it an "Opus-class model." Independent benchmarker Artificial Analysis scored it 54 on its Intelligence Index, placing it fourth among frontier models, and measured its hallucination rate rising from 25 percent to 54 percent versus the prior version. 

Why it matters: The model got more accurate and more confidently wrong at the same time, which is the trade-off nobody puts in a launch post. Treat vendor benchmarks as marketing and go find the independent scoreboard before you trust a tool with work that has to be right.

4) OpenEvidence launches a tool that grades the evidence behind its own answers

Summary: On July 9, OpenEvidence, a clinical AI search tool used by more than 915,000 US clinicians, launched EvidenceGrade, a feature that scores and displays how strong the underlying medical evidence is behind each answer, using the same GRADE framework relied on by the World Health Organization and Cochrane. The company also announced a rollout across every NewYork-Presbyterian hospital and care site, plus Columbia and Weill Cornell Medicine. 

Why it matters: Most AI tools hand you a confident answer and tell you nothing about how solid the source underneath it is. Showing the strength of the evidence is becoming a product feature rather than a nicety, and it is a fair bar to hold any AI tool to before you act on what it says.

5) Illinois signs the strongest state AI safety law in the country

Summary: Governor Pritzker signed the AI Safety Measures Act on July 6, requiring developers with more than $500 million in revenue to publish catastrophic-risk plans, report serious safety incidents to the state within 72 hours, and submit to annual independent third-party audits, a first in any state law. It includes whistleblower protections and takes effect in 2028. 

Why it matters: With California and New York already in place, three states now cover roughly 40 percent of the US AI market, which effectively sets a national standard without Congress. Independent third-party audits are the piece to watch, because that is how safety claims stop being self-reported.

Days before one of enterprise software's loudest voices went on television to call renting your intelligence a mistake, a public company had already moved its money. No press release, no keynote, just a founder's post about a change in how it operates.

In late June, Brian Armstrong posted that Coinbase had cut its AI spend nearly in half while usage kept growing. Coinbase rebuilt how it consumes AI, making cheaper open-weight models the default through its own gateway, caching heavily, and routing each task to the cheapest capable model. The model switch was only one of five changes. Much of the saving came from caching, which just means reusing answers the company had already paid for instead of paying to generate them again. The reuse rate jumped from about 5 percent to 60 percent.

The loudest thing about the move was the invoice. The company had simply decided its intelligence was worth owning rather than renting in full.

You Cannot Outsource the Thing Your Business Runs On

Every public company answers to shareholders and a board, with a duty to run the business predictably. Leaders can tell you, nearly to the dollar, what next year will cost in salaries and rent. AI is becoming as central as either, and it is the only core input where the price, the supply, and the supplier's intentions all sit outside the company's control.

Rent all of your intelligence from one lab, and the first two are gone on day one. The lab decides what you pay, and can change it with a new model or a new pricing plan. It decides what stays available, and can retire the model your workflows were built on. And when a government steps in, the decision is not even the lab's. You already watched one switch off the most powerful model on the planet on a Friday night, for everyone at once.

The supplier's intentions are the part nobody prices in. A frontier lab is not a utility, content to sell power and stay out of your business. These are the most ambitious companies on earth, and they are climbing. Autocomplete became coding assistants, and coding assistants are becoming agents that do whole categories of knowledge work end to end. Every release moves them closer to the work their customers bill for, and the rent you pay helps fund the climb. If your industry is profitable, assume they have noticed.

A supplier who can reprice you, outbuild you, or lose the right to serve you overnight is a dependency, whatever the contract calls it. That is the case for owning your intelligence, at least the part that would hurt most to lose.

Good Enough Is a Price, Not an Insult

Before the strategy tripped me, the vocabulary did. Everyone just says open source and that confuses me, because most of these models are not open source the way people think. Open source hands you the recipe. Open weight hands you the finished dish. You never see how it was made, yet the dish is yours to keep, serve, and reheat as often as you like, on machines you choose, for nothing beyond the cost of running them. That is the kind of model Coinbase now defaults to.

Once the dish is free to own, the buying question changes. Good enough becomes a price, the discipline of paying frontier rates only for frontier work. GLM 5.2 lists around $1.40 per million input tokens, the unit AI usage is billed in. A frontier model runs closer to $5 for the same work, more than three times the price, while the cheaper model scores competitively on the benchmarks that matter. Across millions of calls a month, the frontier premium becomes a tax on work that never needed it.

GLM comes with an asterisk. Several of the strongest open-weight models come from Chinese labs, and US lawmakers are questioning how much to depend on them. Even Microsoft is navigating that tension in plain view. In early July, GitHub made a Chinese-origin open-weight model a lower-cost option inside Copilot, and Microsoft is weighing a self-hosted open model for parts of Copilot Cowork. The honest move is neither to wave that away nor to rush toward those models. Weights you download run where you put them and report back to no one. Who built the model and who controls it are two different questions, and the second one can always be you.

After the model choice comes routing, deciding which task goes to which model, the way Coinbase's gateway sends every job to the cheapest model that clears the bar. Skip that step and the bill finds you anyway. Uber burned through its whole 2026 AI coding budget in four months and now caps engineers at about $1,500 per tool a month. Every finance team is about to run the same experiment. Uber ran it first and learned, in real dollars, that making frontier the default does not survive a budget review.

Your Data Is the Alpha, and You Are Giving It Away

A cheaper bill is only half the argument, and it is the smaller half. The other half is what leaves the building with every request you send to a rented model.

Alex Karp, the CEO of Palantir, was the voice on television. He put it bluntly on CNBC in early July. Pay a frontier lab and you buy tokens while handing over your proprietary data, your workflows, and the judgment underneath them, the things that make your business worth paying for. The customers who understand this, he argued, want to "own the means of production." Take Karp with a grain of salt, since Palantir sells the sovereignty layer he says you need, and Forbes called the interview a sales pitch. The worry is still real, and it does not vanish because the person naming it has something to sell.

In my own industry, the stakes are high. After a decade-plus in healthcare, I know the rule that never bends. Patient data cannot wander. Protected health information, the identifying medical data the law restricts, cannot leave the systems cleared to hold it. That is why regulated organizations, in banking and telecom as much as healthcare, now run open-weight models on their own hardware. One university hospital in Europe has already published the receipt. It wired an open-weight model into its Epic records system, the software that holds the patient chart, ran it entirely inside the hospital's own network, and had more than a thousand clinicians using it within five months. For that hospital, the cheaper bill is beside the point. Owning the deployment is the only version the law allows.

Outside the regulated world, though, the honest numbers say the switch has not arrived. Open-source models' share of enterprise production usage actually fell last year, to 11 percent from 19, by Menlo Ventures' count, even as developers reached for open models more than ever. Owning your intelligence is not free. Self-hosting means buying the hardware, staffing the deployment, and answering the pager when it breaks. That trade is the consideration playing out right now, company by company, and the outcome is not settled.

Decide What to Own Before You Decide What to Rent

So what does an owner actually do on Monday? Build a portfolio, on purpose. Decide which workloads are hard enough to need the frontier, and route the rest to a model you own and can run without watching a meter. IDC projects that by 2028, 70 percent of leading AI-driven enterprises will route across several models rather than depend on one. The mature version of this is not a switch from renting to owning. It is knowing which is which.

I am in the middle of this myself, slowly building a stock research agent for my own use. The decision I keep circling is whether to run it on a model I own or a frontier API I rent. I want to control the costs and only tap into a frontier model when real inference and intelligence is needed. The project is slow, my time going to house matters more than code, and I have not made the call. I am evaluating these decisions in real time, separate from the news cycle. Not everything we share needs to be a finished product. Sharing the deciding is the more useful thing anyway. My own small agent and a hospital data center come down to the same question, own it or rent it.

You do not need to have made the switch to think like an owner. Coinbase changed how it buys, and I am still deciding. What changed was the question we walk in with, from which model is best to which parts are worth owning and which are fine to rent. Which models exist, what they cost, and whether they stay switched on are decided far from your desk. What you own, and what you rent, stays on your side of the screen, no matter how capable the tools get. Which of your workloads truly need the frontier, and which just need a model you control and can run a million times?

The question is the first thing worth owning.

The Assumption

The name says code. The tool arrived in the terminal, the plain text window where programmers type commands. For most professionals, that settles the question, and the conclusion is a fair one. Nothing on the surface of Claude Code invites an operations manager or a program lead to look closer.

Where It Breaks Down

The developers who adopted it refused to keep it in its lane. They pointed it at cluttered folders, at inboxes, at documents with no connection to software. Anthropic watched that happen and drew the obvious conclusion. Rather than steer those users back to the terminal, the company removed the terminal and shipped the same engine as an ordinary app called Claude Cowork. Then it published data on what people actually do with it. Across 1.2 million sessions sampled in late May, software development accounted for 8.7 percent of the work. The largest category, at 33.4 percent, was business process work, the reconciling and report building that fills an ordinary week. Anthropic has a name for that territory, the work around the work.

What Is Actually Going On Here

Set the name aside, and Claude Code is the agent this newsletter built across Vol 24 through 27. Vol 24 named its three parts. A reasoning loop that finishes a step, judges the result, and chooses its next move without waiting for you. Tools it can call to act on the world, such as file access or web search. Memory that carries what it learned in step three into step nine. Anthropic's own documentation describes the behavior in those terms, saying it "plans the approach, writes the code across multiple files, and verifies it works."

Read that line again with the word code taken out. A system that plans, works across files, and checks its own result is a system for finishing a job. Code was only the first material anyone handed it. Cowork is that same machine handed a different material, a point Anthropic makes plainly, describing it as "the same agentic architecture that powers Claude Code, with no terminal required."

The Dial You Actually Operate

Handing work to an agent raises one practical question. How far does it get to run before it asks you? Cowork answers with three settings, and choosing well among them is the skill that travels everywhere.

• Manually approve. Claude stops and asks before each action, and you allow or deny it. Slow, and correct for anything you cannot undo.
• Automatically approve. Claude keeps moving and screens its own actions as it goes, including a check for prompt injection, the trick where a rigged file or web page smuggles in instructions you never gave. Claude blocks what it judges unsafe. The screening costs something, so this mode consumes more of your usage allowance.
• Skip all approvals. No pauses and no checks. Reserve it for tasks where you already trust every file and tool involved.

Anthropic states the ceiling plainly. No safeguard is perfect, and none of them stand in for your judgment. When money, sent messages, or irreplaceable files are in play, stay close to the work.

From here on, every agent you meet will have some version of this dial, named or buried. Find it before you hand over anything that matters.

The Revised Mental Model

The dividing line was never code and everything else. It runs between work you can define and verify, and work you cannot.

Which puts the weight on the skill Vol 24 already named. Goal definition. Where a chatbot waits for your next instruction, an agent takes your definition of done and runs at it. A loose goal yields a wandering agent, and a sharp one yields work you can use. Brief it the way you would brief a capable new colleague. Name the outcome, point to the right materials, say what finished looks like, then read what comes back with real attention.

Where It Still Breaks

An agent with permission to act on your files can act wrongly on them. Cowork requires explicit approval before it permanently deletes anything, a real guardrail, though it does not cover every way a task goes sideways. Vol 26 explained the deeper failure: memory ages and context drifts across a long run, so an early misread compounds quietly instead of getting caught.

Before you try it, know one honest limit. What Claude remembers about you in ordinary chat does not yet carry into a Cowork session. Inside Cowork, memory lives in projects, the same walled workspaces from Vol 40. Cowork runs on paid plans, and the web and mobile versions remain in beta, reaching Max subscribers first.

What to Watch For

• A goal you would not accept from a new hire is a goal an agent will botch. Vague in, wandering out.
• Long autonomous runs deserve more scrutiny than short ones, not less, because drift accumulates where nobody is watching.
• Any vendor selling you an agent should show you its approval controls in under a minute. A slow answer there tells you something.
• Files an agent can reach are files an agent can change. Grant the narrowest access the task actually needs.

How This Connects

The Claude Deep Dive closes here. Vol 39 was first contact on the free plan. Vol 40 opened up memory and projects, where context accumulates. Vol 41 put that context to work inside Artifacts on real deliverables. Today the platform reached its edge, where the tool stops answering questions and begins finishing jobs. Four parts, one honest account of what Claude is actually like to use, limits included. The agent you met today is the one the Agents series built in Vol 24 through 27, now shipping inside products you can open this afternoon. Next week, a Flashback on embeddings, the quiet idea from Vol 5 running underneath Claude's memory, Copilot's search, and nearly every tool you touched this year.

Part 4 of 4 in the Claude Deep Dive series.

Your AI Stack Has Freeloaders

In ten minutes you can know which AI tools are earning their place in your week and which ones are quietly billing you for the privilege of being ignored. That clarity is cheap to get, and almost no one has it.

You collected these tools honestly. A subscription after a demo that impressed you, a trial that never got canceled, a transcription app a coworker swore by. Each one made sense the week you signed up. None has been asked to justify itself since.

The move that fixes this is an adversarial audit. You assemble the evidence, then make an AI argue the case for cutting every tool you want to defend. Claude or ChatGPT runs the analysis and builds the argument against your favorites. The cancellations stay yours. Why this matters: an AI stack that compounds gets built by subtraction as often as by addition.

The Workflow

1. Build the ledger (2 minutes)

Open your billing page and list every AI tool you pay for, plus the free ones you depend on. For each, write the monthly cost, the job you bought it for, and the last real thing you used it for. Everything downstream runs on this list, so guess honestly.

2. Get the ROI read (3 minutes)

Paste the ledger in and ask for the analysis you have been putting off.

Copy/Paste Prompt: "Act as a blunt operations analyst. My AI tool ledger follows. [PASTE YOUR LEDGER. For each tool, give the name, the monthly cost, the job you bought it for, the last real thing you used it for, and roughly how many times you opened it in the last 30 days.] For each tool, tell me the job it is actually doing today, whether another tool on this list already does that job, and whether the time it saves me plausibly beats what I pay plus the friction of keeping it in my workflow. Then sort every tool into keep, cut, or consolidate, and name the single cancellation that would cost me the least and save me the most. Ask me up to two questions first if my ledger is missing something you need."

3. Let it argue against you (2 minutes)

The tools you most want to keep are the ones you have stopped examining. Hand those over.

Copy/Paste Prompt: "Now argue against me. I want to keep [LIST THE TOOLS YOU ARE DEFENDING]. For each one, make the strongest possible case that I should cancel it today. Use my own ledger against me. Name the free or cheaper alternative that covers most of the job, the sunk-cost story I am probably telling myself, and the specific evidence I would need to produce to justify keeping it. Do not soften it."

4. Build the rubric you keep (2 minutes)

Before you close the chat, get something reusable out of it.

Copy/Paste Prompt: "Turn this session into a one-page tool audit I can run every quarter without you. Give me five questions to ask about any tool I pay for, a simple keep, cut, or consolidate decision rule, and three warning signs that a tool has stopped earning its place. Keep it short enough to fit on one screen."

Ask Claude for it as an Artifact and you get a document you can reopen and edit next quarter.

5. Make one call today (1 minute)

Read the argument against your favorites. The AI does not get to cancel anything; you do. Pick the tool with the weakest defense and cancel it before you close the tab.

The Payoff

You walk away with a keep, cut, and consolidate verdict on every AI tool you pay for, and a one-page rubric you can rerun each quarter in five minutes. Your stack stops growing by accident. Tools have to earn their renewal, and the money you free up funds the one that compounds.

The AI Concept You Just Used

This is the adversarial audit. The instinct with AI is to ask it to build your case, and it will do that all day. Asking it to dismantle your case is the harder request and the more valuable one. The move works anywhere you have stopped questioning a recurring cost, whether that is a standing meeting, a weekly report nobody reads, or a vendor contract that renews itself.

Transparency & Notes

• Free tiers cover this. Claude, ChatGPT, and Gemini all handle the analysis, and Artifacts work on Claude's free plan once Code execution and file creation is switched on in Settings.
• Keep employer billing data and negotiated vendor pricing out of the chat. Round the numbers if you need to.
• The AI cannot see your usage logs. It knows only what you tell it, so a flattering ledger produces a flattering result.
• Cost is not the only measure. A tool you open twice a quarter for something high stakes still earns its place. Let the rubric say so.