You're Not In The Hamburger Business

90.9% — Claude Opus 4.7's score on Harvey's BigLaw Bench, the benchmark Harvey itself publishes to measure when an LLM substitutes for billable-hour work. Harvey wrote the test. Anthropic just topped it. Same week, Anthropic wrapped Harvey inside Claude as a callable plug-in and shipped twelve practice-area workflows for legal. The company that owns the benchmark now owns the test, the workflow, and the meter on every lawyer-hour that runs through it. Ray Kroc would have understood the architecture. The McDonald brothers wouldn't. You're not in the burger business.

Run the Tunguz number on your AI bill this week. Tomasz Tunguz published the math today: state-of-the-art AI email costs $22 to $130 a month per user on Claude or OpenAI. The same workflow on a smaller model costs ten to twenty times less. Running locally on the user's own GPU costs zero. If you're paying for a frontier model where a smaller model would do, you're buying retail labor in a market that just commoditized. Pilot one local-inference workflow this week (documentation summarization, inbox triage, meeting-note prep). Measure the gap. If it's under 20% on quality, you've found a defection point worth real money.

Audit your highest-priced consulting line item. Alberta's Ministry of Infrastructure killed a $54 million consulting RFP and stood up a six-person internal team with Claude, Copilot, Gemini, and Lovable. Ten months later: both systems live, 643 users in production, $2.64 million total cost projected. 95% reduction. The specific moment AI did the work the contractors used to charge for: 50 hours of staff screen-recordings pumped through Gemini's vision API at one cent per image, producing structured build specs in minutes. Find your $54 million line. Stand up a three-person internal team for ten weeks. Either you validate the spend or you find 95% in the seat cushions.

Re-underwrite your SaaS pricing model around the agent-hour, not the seat. Mintlify killed seat-based pricing today with the line "seat-based SaaS has no future where agents are the primary users." Anthropic raised the meter on its own subscriptions the same morning. Both moves admit the same thing: the unit of value is the agent-hour, not the chair. Take your top 20 accounts. Look at per-user API call volume from twelve months ago versus last month. If the multiplier is 30x or more (it almost always is now), your gross margin is bleeding faster than your finance team has flagged. The board conversation is coming. Initiate it.

Keeping up with AI is hard. We know — we do this daily. If any of our action items strike a chord, or you simply want an Outsider perspective on how you can improve your business with AI, we'd love to hear from you.

Outsider Labs Advisory →

Thursday, May 14

1. Anthropic booked Harvey at 90.9% on Harvey's own benchmark. Claude Opus 4.7 hit 90.9% on Harvey's BigLaw Bench, the legal industry's most-watched AI substitution benchmark. Two days earlier, Anthropic shipped Claude for Legal with twelve practice-area plug-ins, twenty integrations including Westlaw, and Harvey itself wrapped inside Claude as a callable agent. Harvey's statement: "Anthropic remains a critical partner." Read the language. They became a tenant. (Full analysis below.)

2. Mintlify killed seat-based pricing. Three frontier labs shipped agent-CLIs the same day. Mintlify's one-line announcement: "seat-based SaaS has no future where agents are the primary users." Same morning, Anthropic shipped Claude Code 2.1.142 with multi-agent flags, OpenAI put Codex on mobile with HIPAA, xAI launched Grok Build to SuperGrok Heavy. The pricing model is breaking in both directions because agents are the customers now. (Full analysis below.)

3. The US cleared H200 chips for Alibaba, ByteDance, and Tencent. Approximately ten Chinese firms got approval, including the three giants that anchor the entire Chinese AI stack. Trump's Beijing delegation included Jensen Huang and Elon Musk. Nvidia hit $5.5 trillion market cap as Huang's plane landed. Same morning, Anthropic published a white paper arguing the chip supply chain is the entire policy lever and committing to AGI by 2028. Two American labs took opposite positions on China policy on the same day.

4. Alberta replaced a $54 million consulting bid with $2.64 million in public servants. Three previous attempts to replace two legacy government systems had failed over a decade. Big Four consultants bid $54 million for one of the two systems on a four-year timeline. Alberta's Deputy Ministers killed the procurement and stood up a six-person internal team with Claude, Copilot, Gemini, and Lovable. Ten months later: both systems live, 643 staff using them daily, $858,000 spent. 95% cost reduction. The consultancies didn't lose to a competitor. They lost to the customer.

5. Princeton killed its 133-year honor code in a near-unanimous vote. Faculty leaving the room. Students self-reporting. The system worked when cheating was rare. The cited reason for abandoning it: AI cheating is now "generalized." The institution that taught half of Wall Street to trust the gentleman's word just admitted the gentleman is using Claude. Every enterprise CIO should be asking the same question Princeton just answered.

The Real Estate Business

Today, Anthropic was officially declared the enterprise leader. Ramp's May AI Index put paid business adoption at 34.4% for Anthropic versus 32.3% for OpenAI. Four times Anthropic's share a year ago. OpenAI grew 0.3% in the same window. Every newsletter today led with the data flip.

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

Same morning, Mintlify killed seat-based pricing with one tweet: "seat-based SaaS has no future where agents are the primary users." Anthropic, OpenAI, and xAI all shipped coding-agent CLI updates the same day. Two days before, Anthropic launched Claude for Legal, wrapped Harvey inside Claude as a callable plug-in, and topped Harvey's own BigLaw Bench at 90.9%.

Read these as separate items and you get a busy news cycle. Read them as one trade and you get the day the SaaS business model broke in both directions because the agent is the labor. Anthropic raised the floor on the supply side. Mintlify dropped the seat on the demand side. Harvey became a tenant in the building Harvey thought it was building.

We've seen this movie before. Ray Kroc didn't make $34 billion selling hamburgers. He made it on the land under the franchises. "You don't seem to realize what business you're in," Harry Sonneborn told him on a bench with a flashlight on a balance sheet. "You're not in the burger business. You're in the real estate business." The McDonald 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.

That is exactly today's architecture, run forward into synthetic labor. Anthropic owns the benchmark, the test, the model, the workflow, and the toll on every lawyer-hour, every line of code, every CI pipeline that runs through Claude. Harvey runs on Claude. Harvey is also accessible inside Claude. Harvey's customers can now sign up for Claude directly and get twelve pre-configured legal workflows that Anthropic built, not Harvey. The benchmark Harvey designed to validate its own market position now validates Anthropic's position over Harvey. Every Claude for Legal workflow Anthropic ships next, every plug-in added to the vertical, every model upgrade that nudges the BigLaw Bench score higher makes Harvey's standalone product less essential.

The only American lab betting on a different game is the one not in this fight. Apple owns the silicon, the device, and increasingly the local-inference story. Tomasz Tunguz priced the alternative today. State-of-the-art AI email costs $22 to $130 a month per user on Claude or OpenAI. Smaller models cut that ten to twentyfold. Local inference on a Mac mini costs zero. Every dollar Anthropic charges the meter, the M5 chip gets cheaper to run. The first wave of white-collar arbitrage went to Bangalore. The second wave is going to the desk.

If you're building a vertical AI startup on top of a frontier model, you have eighteen months before the lab enters your vertical. "Build a better agent" isn't the defense. Own a workflow the lab can't see from outside the customer's building. Harvey has a real forward-deployed engineering bench. Watch whether it's enough. Everybody else should be writing the same memo to their board tonight.

The model layer is the burger. The labor market is the real estate. The lab that owns the labor market makes its money the way Kroc did. We are not betting on the franchisees.

At COAI today: Full Signal/Noise briefing with the four-story argument is live at getcoai.com.

— 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.

Sources: