Show Me the Money

THE NUMBER: 16.1% — the share of 240 real freelance projects, $140,000 of actual paid human invoices, that Claude Fable 5 now completes to a client-acceptance standard on the Center for AI Safety’s Remote Labor Index. When the index launched eight months ago, the best model on earth managed 2.5%. The grading question isn’t “can it reason.” It’s “would a paying client accept the deliverable.” Hold that number, and hold the slope underneath it. The whole issue hangs off both.

---

There’s a scene everybody remembers from Jerry Maguire, and almost everybody remembers it wrong. Rod Tidwell, on the phone, making his agent scream “show me the money” until the neighbors can hear it. People file it away as a scene about greed. It isn’t. Rod never asks Jerry to work harder, take more meetings, or send longer memos. He’s renegotiating the unit of account. Stop telling me about the effort, the hustle, the activity. Pay me on what actually lands. The kwan, as Rod puts it — the whole package. In 1996 that was a joke about sports agents. This week it became the most important pricing argument in the AI business, and it came from three directions at once.

Direction one: the Center for AI Safety published new Remote Labor Index results, and Claude Fable 5 — the model Washington kept in a locked room for eighteen days — walked out and posted 16.1% on the one benchmark that grades AI the way a client grades a freelancer. Up from 2.5% at launch. Direction two: Alex Karp went on television and asked why anyone is paying for tokens at all — “If it was so valuable… wouldn’t I say I’ll make you $1 billion and I want 30 percent?” Direction three: the labor market kept booking the story backwards. Companies pinned 40% of May layoffs on AI, up from 4.5% across all of 2025, in a month payrolls grew by 172,000 — blaming a machine that currently finishes one job in six.

Three stories, one thesis. The unit of account in this industry is changing from effort to accepted outcomes. The benchmark now measures it. The loudest CEO in enterprise software wants to bill on it. And the companies using AI as a layoff alibi are about to find out the difference between the excuse and the number. Show me the money, indeed.

🧠 The Benchmark With an Invoice Attached

Most AI benchmarks are exams. The Remote Labor Index is a job.

CAIS and Scale AI built it from 240 real projects pulled off the freelance economy game builds, product design, architecture drawings, data analysis, video animation. Twenty-three domains. The actual client briefs, the actual files, over 6,000 hours of human labor that real clients paid $140,000 for, some single projects running past $10,000 and 100 hours. The AI gets the same brief the human got. Then reviewers ask the only question that has ever mattered commercially: would a reasonable client accept this work as delivered?

That grading standard is why the scores looked so bad for so long. A full project isn’t a reasoning puzzle. It’s planning, file handling, quality control, visual consistency, domain judgment, and final packaging — the boring last mile where actual money changes hands. When the index launched last October, the paper’s own language was that frontier agents sat “near the floor”: about 2.5% for the best of them. The market’s smartest machines couldn’t finish one paid job in forty.

Eight months later, Fable 5 finishes one in six. And notice which model it is. Not the cheap one, not the chatty one — the one the Commerce Department pulled off the market June 12 with a private letter, the one that came back this week wearing a government-tested classifier. The model deemed too capable to circulate is precisely the one best at doing paid human work. Nobody at Commerce will put it that way, but the export-control episode and the RLI score are the same finding, filed by two different bureaucracies.

Yesterday we wrote that the market stopped paying for software that says it has AI and started paying for software that gets work finished. We didn’t know CAIS was about to publish the meter. 16.1% is the acceptance rate of the machine economy, and it now updates like a stock price.

Why this matters: every prior benchmark could be dismissed as academic — SAT scores for robots. This one is denominated in invoices. When it moves, it’s telling you what share of the freelance economy just became contestable. Watch it like you’d watch a competitor’s price list.

💲 The 30% Question

Now the pricing side, because a measurement without a price is just trivia.

Alex Karp, doing what Alex Karp does, went on TV around the Palantir–Nvidia deal and took the entire frontier pricing model apart in two minutes. Enterprises are “livid,” he said. They’re “paying for tokens that create no value.” And then the question that should be laminated and taped to every procurement desk in America: if this thing is as valuable as the pitch says, why is the vendor charging by the token instead of taking a percentage of the outcome? “If it was so valuable, let’s say I can make you $1 billion tomorrow. Wouldn’t I say I’ll make you $1 billion and I want 30 percent?”

Let’s be straight about the messenger. Karp is a salesman — maybe the best pure salesman on television right now — and this rant is him talking his book with both hands. Palantir sits one step up the food chain from the labs: it assembles models it doesn’t own into outcomes it prices, and the spread between the token cost going in and the outcome price going out is exactly where his margin lives. Of course the assembler wants the world on outcome pricing; he’s the one collecting the spread. He’s also, per his own telling, increasingly dependent on frontier AI he doesn’t control, which makes “stop paying the labs so much” a very convenient sermon. ProMarket ran the same x-ray on Satya Nadella this week — his model-portability “warning” happens to route every dollar toward Azure regardless of which model wins. Everyone at this table is preaching their own position. That’s fine. Poker players do that.

But being long your own argument doesn’t make the argument wrong, and Karp’s question has a hundred and fifty years of commercial history behind it. Maritime salvage has run on “no cure, no pay” since before Lloyd’s of London standardized it — you save the ship, you get a cut of what you saved; you fail, you eat your costs. Contingency lawyers, real estate agents, investment bankers, sports agents: every high-stakes service where the client can’t evaluate effort but can absolutely evaluate outcomes eventually converges on the same structure. Take the risk, take the percentage. The industries that bill by the hour are the ones where nobody can prove the work moved the needle. So when a vendor insists on billing you by the syllable, read it as disclosure. A token price is a confession about expected value, notarized and published.

The labs’ response this week was Sonnet 5 — near-flagship agentic capability at $2 a million tokens, the industry’s answer to what The Deep View called AI sticker shock. Cheaper tokens. Which concedes Karp’s frame entirely: they’re competing on the price of effort while the RLI is quietly building the price list for outcomes.

The action item: run the 30% question on your own vendors. Ask your biggest AI supplier what portion of the contract they’d tie to accepted deliverables. You’re not negotiating yet — you’re doing diligence. The vendor who engages has a product. The vendor who retreats to token math just told you what they think their product is worth, and you should price the renewal accordingly.

📉 Taking the Blame

The third direction is the labor market, and it’s booking this story exactly backwards.

Implicator flagged the number: 40% of May layoffs were attributed to AI by the companies making them, up from 4.5% across all of 2025 — in a month when U.S. payrolls grew by 172,000. The Wall Street Journal surveyed top economists on AI job losses and got a three-way split, which is the polite academic way of saying nobody knows. Microsoft is cutting another round while its AI and cloud spend tops $100 billion a year. Ford is rehiring workers it cut when the agents didn’t deliver. And the controlled studies are getting darker: Silicon Canals rounded up research showing the “digital coworker” — the agent with the name, the Slack handle, the seat on the org chart — makes the humans around it measurably worse at their jobs. The marketing says colleague. The data says liability.

Set the two numbers side by side. The machine completes 16% of real freelance work. Companies are blaming it for 40% of their layoffs. The attribution is running two and a half times ahead of the capability. That gap isn’t automation. It’s alibi. “AI took the jobs” is doing for 2026 CEOs what “macro headwinds” did for the class of 2023 — a respectable-sounding cover for cuts the spreadsheet wanted anyway, with the bonus that the board hears “efficiency” instead of “we over-hired.”

We’ve seen this movie before, and we know how it ends. The excuse arrives before the capability, the rehiring happens quietly, and then — this is the part people forget — the capability eventually shows up for real, and the companies that cried wolf have no muscle memory for the actual wolf.

Read it this way: if you’re cutting for spreadsheet reasons, own the spreadsheet. If you’re genuinely betting on automation, the move isn’t fewer people — it’s different people. Somebody has to write briefs tight enough for a machine to execute, and somebody has to grade the output at a client-acceptance standard. The acceptance layer is a department now. Staff it before your competitors name it.

⚡ The Slope Is the Story

So is 16.1% scary? Today, honestly, no. The machine fails five of every six jobs it’s handed. Your business is not about to be run by an agent that can’t finish an animation project.

But levels are for tourists. The slope is 2.5 to 16.1 in eight months — six and a half times — and everything underneath that slope compounds. The models get faster. They get better. They get more agentic, which on this benchmark matters most of all, because RLI failures are planning and packaging failures, exactly the deficits agentic scaffolding attacks. Run the line forward eight more months and the lazy linear answer is 64%. I don’t buy the lazy answer, but I don’t buy it from the other side: on work that can be fully specified in a brief, I think you land closer to 80 or 90. The hard tail — the ambiguous, the political, the taste-driven — holds out much longer. The specifiable middle of the freelance economy does not.

And while the acceptance rate climbs, the cost side is collapsing. Tomasz Tunguz published a piece this week arguing that most teams have the whole stack backwards: the model should be the last decision, not the first, because with a decent router 70 to 80% of agent traffic runs on local models that cost roughly nothing, and async batch work runs 90% cheaper than real-time. Brian Armstrong says Coinbase cut its AI spend nearly in half while token usage grew. Brian Roemmele has been running the same math on GLM-5.2 on local hardware and telling everyone to sit down before they read it. (A flag we’re glad to plant: Tunguz has been one of the best sources in this business for years and a steady well for our own thinking — when he lands on “own the routing, pick models last” eight days after we published the same call, that’s not a scalp, that’s the smartest guy at the table pulling up a chair on our side.)

Put the two lines on one chart. Acceptance rates compounding up. Token costs compounding down. The gap between what an outcome sells for and what it costs to produce is widening every month — and that spread is the entire margin pool of the next era. That’s what Karp is really pointing at. The 30% question isn’t rhetoric; it’s a land grab for the spread. The labs that keep selling tokens are selling the input into somebody else’s outcome business, which makes them — say it with us, because we said it Sunday — the guys selling electricity.

The signal: whoever prices the outcome owns the margin. Vendor, assembler, or you.

What This Means For You

The unit of account is changing underneath the whole industry. The benchmark now grades acceptance, the sharpest operators want to bill on it, and the cost of producing an acceptable outcome is falling on two curves at once.

Reprice one vendor relationship this week. Ask for an outcome-linked component and treat the answer as disclosure. Confidence takes the trade; token math is a confession.

Build your internal acceptance rate. Five real deliverables, original briefs, agent output, graded accept-or-reject at client standard. That number, re-run monthly, is your true exposure — and your automation roadmap, in order.

Separate the alibi from the automation. Layoffs blamed on a 16% machine are spreadsheet cuts wearing a lab coat. Make your own workforce plan off your measured acceptance rate, not the news cycle’s.

Own the spread, or at least know who does. Route the cheap work down (the Tunguz playbook), reserve frontier tokens for work that earns them, and price your own product on outcomes before a Karp-shaped competitor does it to you.

The kwan was never about the money alone; it was about terms that tell the truth. Price the outcome, not the effort — and treat the vendor who won’t take that trade as having just quoted you their real number.

Three Questions We Think You Should Be Asking Yourself

• If your AI vendor offered you outcome pricing tomorrow, could you even take the trade? Outcome pricing requires you to define and measure the outcome. If you can’t specify what “accepted” means for your own workflows, the problem isn’t the vendor’s pricing model. It’s your measurement.
• Which of your deliverables could be specified in a brief tight enough for a machine to attempt? That’s not a hypothetical — it’s the exact boundary the RLI is mapping at 2× a quarter. The work that can be fully briefed is the work that gets contested first. Inventory yours before someone else does.
• What’s your move when the acceptance rate crosses 50%? Not if — the slope says when. If your plan for that day is the same as your plan for today, you don’t have a plan. You have a hope with a org chart.

---

“Every single enterprise in this country, these people are livid. They are paying for tokens that create no value… Why are they charging for tokens if it’s so valuable?”
— Alex Karp, CEO of Palantir, July 2026

---

— Harry and Anthony

---

Sources

• New Remote Labor Index results — Center for AI Safety (@CAIS), Jul 1, 2026 · remotelabor.ai (240 projects, $140K, 6,000+ hours) · original RLI paper, arXiv, Oct 2025 (~2.5% launch baseline)
• Palantir CEO Alex Karp criticizes OpenAI and Anthropic token pricing — Yahoo Finance, Jul 1, 2026 · viral clip thread (@Ric_RTP, Jul 1)
• Palantir’s Alex Karp Just Called the AI Industry “Effing Insane” — Anthony Batt, CO/AI, Jul 1, 2026
• Satya Nadella’s AI Warning Is a Sales Pitch — ProMarket, Jul 1, 2026
• WSJ economist survey (40% of May layoffs attributed to AI vs 4.5% across 2025; payrolls +172,000) — via Implicator.ai, Jun 30, 2026 · Microsoft job cuts as AI/cloud spend tops $100B — Implicator.ai, Jun 30, 2026
• The agentic AI rollout has a contradiction at its core — Silicon Canals, Jul 1, 2026
• Most AI Work Can Wait — Tomasz Tunguz, Jul 1, 2026 (routing over models; Coinbase spend −50% via Brian Armstrong)
• Brian Roemmele on Sonnet 5 cost vs local GLM-5.2 (@BrianRoemmele, Jun 30, 2026)
• Redeploying Fable 5 — Anthropic, Jun 30, 2026 · Fable 5 export-control lift and classifier details — Shelly Palmer, Jul 1, 2026
• Why Sonnet 5 signals an efficiency shift in AI — The Deep View, Jul 1, 2026
• CO/AI prior calls: Git-R-Done (Jun 30, the market pays for finished work) · Central Casting (Jun 23, own the routing table) · Nobody Ever Got Rich Selling Electricity (Jun 25)
• Jerry Maguire (Cameron Crowe, 1996) · Lloyd’s Open Form, “no cure, no pay” marine salvage — cultural and historical anchors, not this week’s news