SIGNAL / NOISE
Paid on Acceptance
The Center for AI Safety runs a benchmark with no vibes in it. The Remote Labor Index is 240 real freelance projects — actual client briefs, actual files, $140,000 of paid human invoices across 23 domains — and one grading standard: would a reasonable client accept this deliverable? Not "did it sound smart." Did the work get done, to the standard someone pays for. When RLI launched eight months ago, the best model on earth completed 2.5% of the projects. This week Claude Fable 5 posted 16.1%. The model Washington kept in a locked room for eighteen days walked out and topped the leaderboard for doing actual human jobs.
Same week, Alex Karp went on TV and gave that number its pricing model. Enterprises are livid, he said, "paying for tokens that create no value." Then the question that breaks the industry: if I could make you a billion dollars tomorrow, wouldn't I just ask for 30%? Why are they charging for tokens if it's so valuable? Now, Karp is a salesman — one of the best on television — and he is talking his book. Palantir sits one step up the food chain, assembling models it doesn't own into outcomes it prices, and the spread between token cost and outcome price is exactly where his margin lives. Of course he wants the industry billing on results. But being long your own argument doesn't make the argument wrong. A vendor who bills by the syllable is telling you what he actually believes his product is worth.
Meanwhile the labor market is mispricing the same asset from the third side. Companies pinned 40% of May layoffs on AI — up from 4.5% for all of 2025 — in a month payrolls grew by 172,000. Against a machine that completes 16% of real freelance work, that math doesn't close. AI isn't taking the jobs yet. It's taking the blame, cover for cuts the spreadsheet wanted anyway.
So is 16.1% scary? Today, no. It fails five of every six jobs it's handed. But 2.5 to 16.1 in eight months is a slope, not a level, and everything underneath it compounds: the models get faster, they get better, they get more agentic, and the scaffolding around them improves monthly. Run it forward eight more months and I don't think you get the linear answer of 64%. On work that can be fully specified in a brief, I think you get closer to 80 or 90. Which is why the Jerry Maguire scene is the right frame. Rod Tidwell never asked Jerry to work harder. He made him scream the terms out loud until the terms changed.
Price the outcome, not the effort — and treat the vendor who won't take that trade as having just quoted you their real number.
At COAI today: the full Signal/Noise — with the RLI math, the slope, and the Karp breakdown — is live at getcoai.com.
Which of your workflows could be paid on acceptance instead of effort. That's the exercise we run at Outsider Labs. If the 16.1% slope is forcing the question, that's the conversation we're optimized for.
ONE — A NUMBER THAT SUMMARIZES THE DAY
16.1% — the share of 240 real freelance projects ($140,000 of human invoices) that Claude Fable 5 now completes to a client-acceptance standard, per the Center for AI Safety's Remote Labor Index. Eight months ago the best score was 2.5%. The grading question isn't "can it reason" — it's "would a paying client accept the deliverable." Six-fold in eight months. The tool is learning to get paid. Show me the money, indeed.
THREE — ACTIONS TO TAKE TODAY
Ask your biggest AI vendor for an outcome-linked quote today. Karp's 30% question works as a diligence tool: if the product delivers the value in the pitch, some slice of the price can ride on acceptance. Watch the reaction. A vendor who will only sell tokens just told you how confident they are in the deliverable. Size the renewal to the answer, not the demo.
Run your own Remote Labor Index this afternoon. Pull five deliverables your team actually shipped last month, hand an agent the original briefs and files, and grade the output the way CAIS does — accept or reject, client standard. That acceptance rate is your real exposure number. Re-run it monthly and chart the line; the level matters far less than the slope.
Split the layoff story from the automation number in your planning. 40% of May cuts were blamed on AI in a month the economy added 172,000 jobs, and the machine finishes 16% of real work. If you're cutting for spreadsheet reasons, own that. If you're betting on the slope, staff the acceptance layer instead — somebody has to grade the machine's work, and that's a job title now.
FIVE — STORIES TO KEEP YOU INFORMED
Wednesday, July 1
The real-work benchmark just moved 6× — and the "dangerous" model leads it. Fable 5 tops CAIS's Remote Labor Index at 16.1%, up from 2.5% at launch. Graded on client acceptance, not benchmark points. (Full analysis above.)
Karp asks the question the labs can't answer. "Why are they charging for tokens if it's so valuable?" He's selling the alternative, and he's still right. (Full analysis above.)
Meta will sell the compute it swears it doesn't have enough of. Meta Compute plans a Bedrock-style cloud; the stock jumped 9.3% — best day since April — while CoreWeave fell 14%. The market just paid Zuck for admitting the models aren't the business. The rent is. Together AI raised $800M at $8.3B the same day.
Someone poisoned the agent app store. Attackers slipped malicious skills into OpenClaw's ClawHub marketplace — markdown files that inherit every permission your agent already has, per Palo Alto's Unit 42. A skill is code that happens to be written in English. Review it like a dependency, because that's what it is.
Fable 5 came home wearing an ankle monitor. Back globally after eighteen days, with a government-tested cyber classifier that false-positives routine dev work down to Opus 4.8. You order the Ferrari; if the request smells wrong, the valet brings the Civic. The first regulator-negotiated routing rule inside a frontier model.
MARK TO MARKET
Where the cycle caught up to us this week.
Pick your models last. The routing table is the asset; the model is a rental (us, Jun 23) → "Design your system around routing, not around models. Pick your models last" (Tomasz Tunguz, Most AI Work Can Wait, Jul 1). One caveat, gladly stamped: Tunguz is one of the best sources in this business and a steady well for our own thinking — this receipt reads less like a scalp and more like the smartest guy at the table pulling up a chair on our side.
The model is not the moat. The lever is free; value lives in what you own around it (us, Jun 16) → two dozen investors at 500 Global's VC Unlocked converged "on exactly one thing": the model is no longer a moat; "code is disposable" (500 Global, Jul 1).
The tape doesn't lie. We just read it early.
— Harry and Anthony
Sources:
WSJ economist survey: 40% of May layoffs attributed to AI, vs 4.5% across 2025; payrolls +172,000 — via Implicator.ai, Jun 30, 2026
Together AI raises $800M at $8.3B — TechCrunch, Jul 1, 2026