Git-R-Done

"We Have AI" Doesn't Get 'Er Done

Claude Sonnet 5 shipped this morning, and the tell is what it doesn't do. It doesn't want to chat. It makes plans, drives a browser, runs a terminal, and works on its own at a level that took a bigger, pricier model a few months ago — for two dollars a million tokens, set as the default on the free tier. The chatbot you've been paying to talk to just got replaced by a worker you pay to finish things. Ethan Mollick called it the twilight of the chatbots the same afternoon. He's right.

Here's the part that should make you sit up. On the same day, Tomasz Tunguz published the market's verdict on exactly this shift, and it's brutal. Across 87 public software companies, the seat-priced application layer — Salesforce, Workday, ServiceNow, the stuff that runs your company — is down 36% on the year. The plumbing underneath it, the compute and data and security that agents actually run on, is up 68%. Both buckets grew revenue about 21%. Growth didn't separate them. One group can show work getting done. The other can only say it "has AI."

That's the whole issue. "We have AI" is the sound a company makes right before it admits it hasn't shipped anything. The market quit grading the pitch and started grading the output, repricing every line by whether the answer is a demo or a deliverable. Benioff already answered for Salesforce: 9,000 heads cut to 5,000, because Agentforce does the work now.

And watch what it cost Anthropic to get its own best models back. Commerce gated Fable 5 and Mythos 5 eighteen days ago over cyber capability; Mythos was good enough to embarrass the NSA's own defenses. Anthropic didn't win them back with a lobbyist. It shipped Sonnet 5 with the cybersecurity capability deliberately stripped out, and said so in the release notes. It had to show, not tell. The gate came down the same week.

So don't file this under "cheaper Claude." The unit of value moved from access to a model — now a commodity the open-weight crowd sells for pennies — to the work a model finishes. Saying you have AI is table stakes. Showing what it closed is the only thing anyone's paying for now.

Git-R-Done stopped being a punchline this week and became the business model. Show what got finished, or get repriced for talking about it.

At COAI today: the full Signal/Noise — the eighteen-day gate and what really reopened it, the Tunguz sector map, and why Ford is rehiring the people it let go — is live at getcoai.com.

Outsider Labs Advisory →

−36%. That's the one-year haircut the market took out of seat-priced business software — Salesforce, Workday, ServiceNow — even as the compute-and-data plumbing agents run on rose 68% on the same revenue growth. Same growth, opposite verdict. The market stopped paying for software that says it has AI and started paying only for what visibly gets the work done. Sonnet 5 shipped a worker at $2 a million tokens today. The −36% is the repricing of everything it's about to replace.

Sort your software budget into "token path" and "seat." Tunguz's map is blunt: infrastructure, security, and messaging — the things agents run on — are up on the year; per-user application seats are down 36% because agents substitute for them. Pull your SaaS renewal list today and mark each line one or the other. The seats are your renegotiation targets. Go in before the vendor reprices for you.

Route the bulk work to the worker tier, and measure what it finished, not what it burned. Sonnet 5 at $2/$10 is built for the agentic jobs running up your bill; reserve frontier tokens for the work that earns them. Then kill the usage leaderboard — Meta ranked staff by tokens consumed (top user: 281 billion) and Amazon had to shut its version down with a memo: don't use AI for the sake of it. Count outcomes, not tokens.

Dual-source before the gate swings back. Fable and Mythos went dark for eighteen days on a government order, then came back just as fast. The lesson isn't which frontier won. It's that frontier access is a turnstile now, not a wall. Stand up an open-weight model you control for anything you can't afford to have gated on someone's bad mood.

Tuesday, June 30

Sonnet 5 clocks in — the chatbot era is over. Anthropic's most agentic Sonnet yet runs browsers and terminals on its own at near-Opus quality for $2/$10, defaulted to Free and Pro. The interface you used to talk to is now a worker you assign. (Full analysis above.)

The gate becomes a turnstile — Fable 5 and Mythos 5 come back. Commerce lifted the export controls it slapped on Anthropic's two best models eighteen days ago; access returns tomorrow. The price of readmission: shipping Sonnet 5 with its cyber capability stripped out. (Full analysis above.)

China stops hacking your network and starts recruiting your people. CrowdStrike says Chinese entities were more than half of state-sponsored intrusions on tech and AI firms in the year through March, and the vector is shifting to insiders and social engineering. Startups, with no security budget, are the soft target.

The layoffs found the master's degrees. A California Policy Lab study ties high-AI-exposure workers with bachelor's and graduate degrees to a 20% jump in jobless claims, concentrated in the Bay Area. The "AI hits entry-level first" story is backwards. It hits the expensive first.

Amazon shops around on the lab it bankrolled. Reports say Amazon is weighing OpenAI and its own Nova models as Anthropic raises prices, and one state landed Claude at a 50% discount. The company that wrote the check is now the competitor across the table.

Where the tape caught up to us this week.

The keepers of the culture stay — even after you fire them. Talent is a rental, but the integrator who understands the thing in ways the model can't is the real moat (us, The Keeper of the Culture, Jun 22) → Ford quietly rehired workers it had replaced with AI, after finding the models couldn't match their grasp of the car and its safety (The Deep View, Jun 30).

The meter reached the seat. Intelligence deflates toward the cost of electricity, and the value moves above the model to the fulcrum — not the per-seat license (us, The Meter's Running, Jun 3, and Show Me Where to Put the Fulcrum, Jun 16) → the public market marked it in the open: seat-priced business apps down 36% on the year while the infrastructure they run on ran up 68% (Tomasz Tunguz, Jun 30).

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: