You’re Not In The Hamburger Business

THE NUMBER: 90.9% — Claude Opus 4.7’s score on Harvey’s BigLaw Bench, the benchmark Harvey itself publishes to measure how well an LLM substitutes for billable-hour legal work. Harvey wrote the test. Anthropic just topped it. The same week, Anthropic wrapped Harvey inside Claude as a callable plug-in and shipped twelve practice-area workflows for legal. The model that aced the question “can this replace a lawyer?” is now sold by the company that owns the benchmark, the test, the workflow, and the meter. Ray Kroc would have understood the architecture. The McDonald brothers wouldn’t.

Today, Anthropic was officially declared the enterprise leader: Ramp’s May AI Index put Anthropic at 34.4% of paid business adoption versus OpenAI at 32.3%. That’s four times Anthropic’s share a year ago. OpenAI grew 0.3% in the same window. The headline reads like a victory lap.

The actual story landed in the footnotes. Starting June 15, your $20 to $200 monthly Claude subscription splits in two. Full use inside Anthropic’s own apps. A metered pot for everything else (Conductor, Zed, T3 Code, claude -p running in your CI, the Agent SDK called from third-party tools, Claude Code GitHub Actions). When the pot empties, you flip to pay-as-you-go at standard API rates. The April 4 ban on third-party agents like OpenClaw is reversed, and the meter is on.

Same morning, Mintlify killed seat-based pricing with one line: “seat-based SaaS has no future where agents are the primary users.” Three frontier labs (Anthropic, OpenAI, xAI) shipped coding-agent CLI updates the same day. Claude Code 2.1.142 with multi-agent orchestration flags. Codex on mobile with HIPAA, SSH, and Hooks. xAI’s Grok Build, an agentic CLI, opened to SuperGrok Heavy subscribers. Two days before, Anthropic launched Claude for Legal with Harvey itself plugged in as a callable agent, plus Westlaw and Intapp integrations and a Freshfields partnership.

Read in isolation, that’s a busy news cycle. Read together, the SaaS business model just broke in both directions on the same day, because the agent is the labor. Anthropic raised the floor on the supply side. Mintlify dropped the chair on the demand side. The unit of value is no longer the seat or the chat. It’s the agent-hour. And Anthropic just published a white paper putting AGI on a 2028 calendar with explicit “beat the adversaries” geopolitical framing, because owning the labor market is now an industrial policy.

We’ve seen this movie before. Ray Kroc didn’t make money selling hamburgers. He made it on the land under the franchises. The McDonald brothers thought they were running a hamburger business. They were running on someone else’s real estate, and they got bought out for a flat $2.7 million and lost the royalty stream forever. Today is the day Anthropic became Ray Kroc.

Anthropic Bought the BigLaw Benchmark

🦞 On Tuesday, Anthropic shipped Claude for Legal. Twelve practice-area plug-ins. Twenty-plus integrations with the tools BigLaw already runs on (Westlaw, document management, conflict-checkers, time-keeping). An Intapp deal. The Freshfields partnership. And the structural move: Harvey, the legal AI company every BigLaw firm has been piloting for two years, is now callable from inside Claude as a plug-in.

Two days later, Anthropic published the benchmark score. Claude Opus 4.7 hit 90.9% on Harvey’s own BigLaw Bench, the legal industry’s most closely watched AI benchmark, designed by Harvey specifically to measure when an LLM substitutes for billable-hour work. Harvey’s public statement on the integration was careful: “Anthropic remains a critical partner for Harvey and a key input to our product for customers. Their decision to launch a legal plug-in doesn’t change that for us.”

It changes everything for them. Harvey runs on Claude. Harvey is now also accessible from inside Claude. Harvey’s customers can sign up for Claude directly and get twelve workflows pre-configured by Anthropic, not Harvey. Anthropic owns the benchmark, the test, the model, the workflow, and the toll on every lawyer-hour that runs through it. Harvey just became a tenant in a building Harvey thought it was building.

This is the Ray Kroc move, exactly. The McDonald brothers made the burgers. They invented the assembly-line kitchen. They owned the recipe. Kroc owned the land underneath. “You don’t seem to realize what business you’re in,” Harry Sonneborn told Kroc, sitting on a bench in the dark with a flashlight on a balance sheet. “You’re not in the burger business. You’re in the real estate business.” The brothers got a one-time check and a handshake on royalties Kroc later refused to honor. They died watching a company that bore their name pay rent to somebody else.

Harvey hasn’t been bought out. But every Claude for Legal workflow Anthropic ships next, every plug-in added to the legal vertical, every model upgrade that pushes the BigLaw Bench score higher, makes Harvey’s standalone product less essential. The benchmark Harvey built to validate its own market position now validates Anthropic’s position over Harvey.

The strategic read: Vertical AI startups should assume the stack they sit on is rented, not owned. If you’re building a vertical agent on top of a frontier model and the lab enters your vertical, you have roughly eighteen months. “Build a better agent” isn’t the defense. The defense is owning a workflow the lab can’t see from outside the customer’s building (forward-deployed engineering, regulatory permissioning, signed-data partnerships that travel with the customer). Harvey’s bench on that work is significant. Watch whether it’s enough. The rest of the vertical AI category should be writing the same memo to its board tonight.

The Buffet Is Closing. So Are the Chairs.

💲 Starting June 15, your Claude subscription splits in two. Full use of Claude inside Anthropic’s own apps (the website, Claude Code in the terminal, Cowork). For anything called through a third party (Conductor, Zed, Jean, T3 Code, the Agent SDK from a non-Anthropic app, GitHub Actions, claude -p running in CI), you get a metered $20 to $200 monthly pot depending on plan. When it’s empty, you can keep going at standard API rates.

This reverses the April 4 ban Anthropic slapped on after Pro and Max subscribers figured out that wiring OpenClaw into a $200 plan let them burn thousands of dollars in API tokens a month. The April move was a panic. The June move is the considered one. Anthropic is admitting in the most expensive way possible that subscription power users were upside down for a year, and the company is finally pricing the labor like labor.

Same morning, Mintlify (the documentation platform every developer-tools company uses) eliminated seat-based pricing entirely. The reason fit in a tweet: “seat-based SaaS has no future where agents are the primary users.” Mintlify’s customer is now an agent (Claude, Cursor, Codex) reading documentation to generate code. You can’t charge by the seat when the seat is an LLM hitting your servers a hundred times more than a human reader would. The math collapses.

Both moves are the same trade run from opposite sides of the contract. Anthropic raises the meter on the supply side. Mintlify removes the meter on the demand side. The thing they’re both admitting: the seat is dead. The chat is dead. The unit of value is the agent-hour, and the SaaS pricing playbook every software company has used since 2008 is suddenly an open question.

Three frontier labs shipped coding-agent CLI updates the same day. OpenAI put Codex on mobile and added SSH, Hooks, Access Tokens, and HIPAA compliance. xAI shipped Grok Build (an agentic CLI) to SuperGrok Heavy subscribers. Anthropic dropped Claude Code 2.1.142 with new multi-agent orchestration flags. Three companies, one product category, one calendar day. None of them is competing on a chatbot. They’re competing on whose synthetic labor force you rent by the hour.

Why this matters: If you run a SaaS company, your pricing model is a board-level question this quarter. Take your top 20 accounts. Look at the per-user API call volume from twelve months ago, then from last month. If the multiplier is 30x or more (it usually is now), your COGS per seat is wrong by an order of magnitude and your gross margin is bleeding faster than your finance team has flagged. Mintlify just gave you the public language for the answer. The investors will start asking by Q3 earnings. Don’t wait for the call.

The Mac Mini Defense

🧠 Tomasz Tunguz published the math today. State-of-the-art AI email running on Claude or OpenAI costs $22 to $130 a month per user. Take the middle case at $26 raw. A software company seeking a 75% gross margin would charge $350 to $500 per year for that product. Google Enterprise is $11 to $18 a month. A fully agentic email assistant costs roughly twice as much as the email itself.

The same workflow running on a smaller model (DeepSeek V4 Flash, the five SLMs Anthropic has been benchmarking for tool-calling, the free DeepSeek V4 Flash NousResearch just put up) costs ten to twenty times less. The same workflow running locally, on the user’s own GPU, costs zero. Tunguz’s exact line: “the only [reason that matters] is latency.” He wrote that one on Sunday. Today he priced the trade.

NousResearch shipped Hermes today. Persistent local AI agent. Runs on consumer NVIDIA RTX cards and DGX Spark, no cloud dependency. The /goal command lets you assign outcomes instead of managing individual prompts. Same week, somebody trained 240 million fine-tuning examples for $78 in 72 hours and captioned it “the data moat is gone.” DeepSeek V4 Flash matches V4-Pro at 90% lower cost than GPT 5.4 Mini. The Ring-2.6-1T frontier-scale model was open-weighted on Hugging Face today, tested on both American and Chinese hardware. The floor under the labor market keeps dropping.

And Apple is, as Anthony Batt argued this morning, the only major American tech company structurally betting on the defection. (Anthony’s framing on Tim Cook’s succession needs verification before we treat it as event news. The structural point stands either way.) Don’t chase frontier models. Own the silicon. Own the device. Run the labor locally. Let Google, Meta, and Microsoft burn $300 billion fighting for the model layer while you collect the rent on the chip the model runs on.

Read the trade clean. Every dollar Anthropic charges the meter, the Mac mini gets cheaper to run. Every percentage point DeepSeek shaves off the floor, the rent gets harder to enforce. The labor market, like the manufacturing labor market before it, is going to get arbitraged. The first wave of white-collar arbitrage went to Bangalore. The second wave is going to the M5 chip on your desk.

Here’s what to do: Run the Tunguz number on your own AI bill this week. If you’re paying for a frontier model API call where a smaller model would do, you’re buying retail labor where the gig market just commoditized. Pilot a local-inference workflow on something non-critical (internal documentation summarization, meeting-note triage, inbox classification, screen-recording transcription) and measure the quality gap. If it’s under 20%, you’ve found your defection point. The Apple bet is that the average defection point lands lower than $26 a month, and the next eighteen quarters of AI capex decisions will get re-underwritten against that number.

Alberta Killed The Procurement

💲 For more than a decade, Alberta’s Ministry of Infrastructure tried to replace two legacy systems. A building and asset tracker. A construction-project budget tracker. Three previous attempts failed. In 2025 the Ministry ran a formal procurement. Thirteen bids. Four finalists, all major consulting firms. The shortlist price for one of the two systems: $54 million over four years. The other system wasn’t even in scope.

The Deputy Ministers killed the procurement. They stood up a small team of Alberta public servants under what they’re calling the AI Maximalist program. Tools: Claude, Copilot, Gemini, and Lovable. Approach: build working software fast, put it in front of real users, ship updates every two weeks, fix what breaks. Nate Glubish (Alberta’s Minister of Technology and Innovation) published the case study under the title “They Said It Would Cost $54 Million. We Said ‘No Thanks.'”

The numbers, in production. Both systems are live. 643 Infrastructure staff use them every day. The team has spent $858,000 cash over ten months, runs at $118,000 a month, and projects a total cost of $2.64 million to fully deliver both systems. The vendor bid was $54 million for one. A 95% reduction. Both systems instead of one. Four years compressed to ten months. The project lead’s quote: “Folks are legitimately thrilled to be using a new system.”

The specific moment AI did the work the contractors used to charge for: Infrastructure had 50 hours of staff screen-recordings showing how legacy workflows actually moved. Old playbook, McKinsey or Accenture interviews the staff, writes the requirements doc, generates a build spec. Months. Millions. New playbook, pump the video frames through Gemini’s vision API at one cent per image. The model produced structured requirements, user flows, and data models. In minutes. Not days. Not weeks. Minutes.

This is the Lucy thesis at public-sector scale. The May 11 piece on Bessemer’s Ascend case study was a private-sector growth team building a Meta ad funnel with Claude Code slash commands. Alberta is the same shape, on a different problem, at a different scale, with one variable changed. The customer is the same organization that used to write the procurement RFP. When the cost of internal capacity collapses, the org chart stops being a procurement organization and becomes a delivery organization. The consultancies didn’t lose to a competitor. They lost to the customer.

(One date caveat in the spirit of the recent Karpathy/Dwarkesh trap. Glubish’s piece carries an April 6 byline inside the post but is being amplified through May 14 channels. The numbers don’t depend on the timing. The story works whether the announcement broke six weeks ago or this morning.)

The action item: Every CEO and CFO has a $54 million line item somewhere. Look at the budget. Find the highest-priced multi-year consulting contract, ERP rollout, or enterprise software migration currently open. Stand up a three-person internal team with frontier model access for ten weeks and ship a working prototype of the same deliverable. Worst case, you validate the spend with real data. Best case, you discover that 95% of the bill was paying contractors to read your own employees’ screen recordings. Either answer changes the next budget cycle.

What This Means For You

The labor market for synthetic workers got built in front of us today. Anthropic is the staffing agency. Mintlify is the customer who gave up trying to count the workers. Harvey is the vertical hiring hall that just became a tenant. Apple is the contractor who decided to own the building. Alberta is the customer who fired all of them and did the work itself.

Pick a side of the rent. You’re selling labor (an AI-native vertical agent), buying labor (paying the meter to a frontier lab), or building your own (local inference, internal team, the Alberta model). The fence in the middle doesn’t exist anymore. If your model still assumes the seat or the chat is the unit of value, the unit economics are wrong.

Audit your highest-priced consulting line. Whatever your enterprise pays McKinsey or BCG or Accenture for, run an Alberta-style ten-week internal pilot with frontier model access. Either you validate the spend with data the consultants will never give you, or you find 95% in the seat cushions. Both answers are useful. The third option (do nothing) costs you whichever number is bigger.

Watch the SaaS repricing. Mintlify didn’t kill seat pricing because they’re generous. They killed it because they had to. The next quarter of public-software earnings will have a new analyst question on every call: “What’s your per-agent ARPU, and what’s your defense when your customer is a Claude session?” If you can’t answer that on slide one of your next board deck, the deck is wrong.

The model layer is the burger. The labor market is the real estate. Ray Kroc didn’t make $34 billion selling hamburgers. He made it on the land underneath. The lab that owns the labor market makes its money the same way. The companies betting on local inference (Apple, Tunguz’s stack, Hermes on RTX, every team running DeepSeek on a Mac mini) are betting the tenants eventually buy their own land. We are.

Three Questions We Think You Should Be Asking Yourself

Is my pricing model still counting humans when my customer is sending agents?

If you’re a SaaS founder or a product CFO, run the agent-load number on your top accounts. A single human user is now driving 50x to 100x the API call volume of twelve months ago because their agents do the work. Your COGS per seat is wrong by an order of magnitude, and your gross margin is bleeding faster than your finance team has flagged. Mintlify gave you the language. The board conversation is coming whether you initiate it or not.

Am I a tenant in someone else’s labor market, or am I building my own?

Every vertical AI startup running on a frontier model is a tenant. The rent is reasonable today, metered tomorrow, renegotiated the day the lab enters your vertical. Harvey saw it and built a deep forward-deployed engineering bench as the defense. What’s yours? If the answer is “we have better prompts,” you don’t have a moat. You have a feature, and a landlord.

If Alberta cut $54 million to $2.64 million, what’s my $54 million line item?

Look at the budget. Find the highest-priced multi-year consulting contract or enterprise software project on the books. Stand up a three-person internal team with Claude, Gemini, and Lovable for ten weeks. If they can ship a working prototype, you’ve found 95% in the seat cushions. If they can’t, you’ve validated the spend with data you didn’t have before. Either way, you’ve changed how you write next year’s budget.

“You don’t seem to realize what business you’re in. You’re not in the burger business. You’re in the real estate business.”

— Harry Sonneborn to Ray Kroc, The Founder (2016)

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

You’re Not In The Hamburger Business

Anthropic Bought the BigLaw Benchmark

The Buffet Is Closing. So Are the Chairs.

The Mac Mini Defense

Alberta Killed The Procurement

What This Means For You

Three Questions We Think You Should Be Asking Yourself

Sources

Past Briefings

Sully: Brace for impact.

I Am Iron Man

Lucy