Canary in a Coal Mine

THE NUMBER: 11 percent. That’s the share of all spending on Anthropic’s tools that flows to Fable 5, the company’s biggest and most expensive model, more than two months after launch, according to spending data Ramp collects from 70,000 companies. For two years the reflex was automatic: a new frontier model ships, corporate buyers default to it, and the priciest tier takes the crown. Fable broke the reflex. The most expensive intelligence on the market is now a rounding error in the bill. Hold that number, because once you understand why the best model money can buy can’t crack a ninth of the spend, you understand the whole trade.

The most expensive model on earth is a rounding error

Start with the fact, because it’s stranger than it looks.

Fable 5 is Anthropic’s flagship. It is the model the company’s roughly two-trillion-dollar IPO — expected as soon as next month, which would be the largest in history — is implicitly priced on. And more than two months after it shipped, it accounts for only about 11 percent of what businesses spend on Anthropic, on Ramp’s read of 70,000 companies. The pattern that held for two years, where the newest and most capable model became the default and everything older got abandoned, simply stopped. The bulk of the work is being done by cheaper, older, good-enough models, and Anthropic’s own less expensive Opus 5 has already passed Fable in business spending since launching in late July.

The people closest to the company are not spinning it. Miles Clements is a partner at Accel, which has put close to a billion dollars into Anthropic, so this is a man talking down the value of his own position. His line to the Financial Times: “Most people don’t need to operate at the frontier.” The stretch when customers reflexively reached for the top model, he added, “was not a durable era.” Breakthroughs still matter, he allowed — they cure diseases and recruit the best researchers — but from here they’re showcases, not the thing most buyers pay for. Ramp’s own lead economist, Ara Kharazian, put the whiplash plainly: on the old trend you’d expect Anthropic to own the market, but because OpenAI’s newest model was so good and Fable underperformed, it’s been the reverse.

Read that as an operator and it’s a repricing of the entire frontier business. The labs have spent tens of billions of dollars a year training bigger models on the theory that the best model commands the best margin. The Ramp data says the best model commands 11 percent. If that holds, the money in AI is not where almost everyone has been pointing it. So where did it go?

Same mine, opposite blueprints

Here’s the frame that makes it click, and it starts with a coincidence of hardware.

A data center is a mine. Scarce power, purpose-built silicon, racks humming in the dark, and out of that rock you pull something valuable. Bitcoin runs on exactly that mine — the same warehouses, the same grid, purpose-built chips instead of general-purpose ones — and so does artificial intelligence. Same physical plant. Opposite blueprints.

Bitcoin was engineered, from the first block, to be scarce. Twenty-one million coins, ever, and the protocol makes the digging harder on a fixed schedule no matter how many miners show up. Worth being precise here, because the causation trips people up: it isn’t the rising cost of mining that drives Bitcoin’s price. The fixed supply drives the price, and the mining difficulty rises in response. Expensive extraction is a symptom of a valuable coin, not the cause of one. Strip away the noise and Bitcoin is a store of value — digital gold — and stores of value are made to be hoarded and to appreciate.

Intelligence was engineered for the precise opposite. There is no cap. You make more of it by pouring in more power and more chips, the models get cheaper by the week, and the only ceiling on supply is how much you’re willing to spend. Where Bitcoin is digital gold, intelligence is digital labor: a flow, not a store. And a flow doesn’t get hoarded. It gets consumed, and it gets repriced to its marginal cost — which, for a token, is falling toward the price of the electricity that produced it. OpenAI’s cheapest tier now undercuts DeepSeek on intelligence per dollar. Sam Altman says the price of intelligence is dropping something like tenfold a year. Gold rises. Labor gets commoditized. Same mine, opposite curves.

Now the part that makes this a vise instead of a slogan. If intelligence is deflating, why is the hardware inflating? Nvidia (NASDAQ: NVDA) is telling customers that AI-server prices climb around 15 percent next year, with memory alone now close to a third of the bill of materials on a two-million-dollar rack. Tomasz Tunguz, the Theory Ventures partner whose data we cite constantly, named the mechanism this week: the AI supply chain is running “the classic bullwhip effect” — first GPUs were scarce, then memory, then CPUs, and now hard disks. The same commoditization that collapses the price of a token is what stampedes the entire industry into the same scarce power and silicon, and that stampede is exactly what bids the shovel up. Cheaper output, dearer input, and they are not two separate stories. They are one mechanism working both ends. Jevons saw the shape of it in 1865 with coal: make the fuel more efficient and you don’t use less of it, you use dramatically more. Make a token cheaper and agentic demand for tokens jumps fourteen-fold in six months, which is precisely what happened, which is precisely what keeps the memory market tight.

We’ve watched this movie before, in a different theater. The late-1990s telecom build laid so much fiber that the price of bandwidth collapsed toward zero. The bandwidth got cheap; the companies that owned nothing but bandwidth went bankrupt; and the value fled to the layers the glut couldn’t touch — the applications on top and the specialized gear underneath. Intelligence is the new bandwidth. The question is which layer you want to be standing on when the price of the commodity in the middle goes to nothing.

The duel: growth on discount versus dollars and margin

You can watch the whole thing play out as a two-horse race, and the scoreboard is rigged to mislead you.

On Ramp’s numbers, OpenAI grew faster than Anthropic among business users last quarter — call it 82 percent to 76 on spend. Read the headline and OpenAI is winning. Read the fine print and the story inverts. OpenAI bought that growth with a discount: GPT-5.6 is priced well below Fable, and the cheaper model is doing the work. Anthropic, meanwhile, is the bigger and the healthier house. Its revenue is running around 65 billion dollars annualized to OpenAI’s north of 40; it booked its first adjusted operating profit in the second quarter and told investors it expects to be profitable again in the third; it has some 6,000 customers spending north of $100,000 a year. OpenAI is not profitable.

So which one is winning? It depends on which scoreboard you read, and they reward opposite behavior. The growth-rate board hands the ribbon to whoever cuts price hardest, and that’s OpenAI, riding the cheaper model into faster adoption. The dollars-and-margin board hands it to whoever owns the stickiest, highest-paying base and actually turns a profit, and that’s Anthropic. The growth-rate lead is a small-base illusion: even at a slightly slower percentage, Anthropic is adding something like six billion dollars of quarterly revenue to OpenAI’s one. The gap in absolute dollars is widening while the gap in growth rate flatters the smaller number.

And underneath both boards sits the fact that should scare every investor in the space: neither company owns a shred of loyalty. Mindstream, reading the same Ramp data, said it cleanest — two rivals, one market, zero loyalty, with businesses “switching providers as fast as labs ship new models.” This is our own thesis from If You Ain’t First, You’re Last two weeks ago, coming back at us near-verbatim: the model tier earns no allegiance. The switching costs the labs assumed would lock customers in have turned out to be a weekend’s work. That’s wonderful for buyers and brutal for anyone underwriting a two-trillion-dollar valuation on the premise that the base is captive.

The one seam that still pays

If intelligence in general is deflating to the cost of electricity, is there any kind of it that holds its price? Yes. Exactly one, and it’s the tell for everything.

The value that doesn’t collapse is the value the machine can grade itself. Coding is the purest case. When a model writes code, the compiler renders an instant, free, objective verdict — pass or fail, it runs or it doesn’t. That verdict is worth money, because it turns a plausible guess into a checkable result, and checkable results are what businesses actually pay to automate. There’s no more loyalty among developers than anywhere else — GPT-5.6’s cheaper coding tier just proved they’ll jump ship for price and quality in an afternoon — but there is real, durable cash on the table, because the work has a grader.

Which is why the most important number in this issue is one that isn’t in the Ramp data yet. The Ramp snapshot runs through July. The newest cheap coding entrant — Grok’s mid-August push, priced around two and six dollars per million tokens, roughly 60 percent under Claude — landed after the window closed, and it points straight at coding, which is Anthropic’s stronghold. So the freshest, most aggressive price cut in the one segment that still pays isn’t reflected in the 11 percent at all. The canary in our lantern is already looking peaky, and it hasn’t yet breathed the gas that’s coming. Whatever the deflation looks like in July’s numbers, August is worse.

This is the same split we drew in The Usual Suspects: the graded half of intelligence — math, engineering, code, anything with an objective referee — stays valuable and defensible because the referee is free and merciless. The commodity half — the generic memo, the passable summary, the undifferentiated chat — races to the floor. If you’re building a business on AI, the entire game is figuring out which half you’re in, and getting yourself onto the side with a grader.

The other canary

There was a second canary in the mine this week, and it’s worth a paragraph because it runs on the same logic.

A fully autonomous drone killed three civilians in Zaporizhzhia — the first such deaths anyone has documented — and the machine that did it was a consumer Nvidia module worth a few hundred dollars, flying with no operator choosing the final target. It happened back on July 6; it only got confirmed and reported this week. Set that next to a quieter statistic: 57 percent of IT leaders say they expect to pull the human out of the decision loop entirely within a year. One canary is economic — the frontier model going silent at 11 percent, warning that the premium trade is thinner on air than it looks. The other is operational — a three-hundred-dollar board making a kill decision, warning that the human is being removed from the loop faster than anyone’s checking whether that’s safe. Different shafts, same instrument. When the canary goes quiet, the mistake is to study the canary. The signal was never about the bird. It’s about the air, and whether you’re still breathing it.

Digital gold, digital labor, and where you stand

Put it together and the week stops being a jumble of headlines and becomes a single instruction.

The frontier model is a commodity, and the market has now said so out loud, in dollars, with a frontier investor and the data source itself co-signing. Accel’s Clements, a billion dollars in, calls the frontier-only era not durable. Tunguz names the bullwhip that inflates the shovel while the gold deflates. Half the newsletters in the cycle ran some version of “the frontier era just ended.” When the people longest on the trade start narrating its top, that’s not noise. That’s the tape.

So stand in the right place. There are three positions in this mine and only one of them is good. You can own the shovel — the depreciating iron at the top of the bullwhip — and watch Nvidia and the memory makers keep the margin while your asset loses a third of its book value a year. You can sell the commodity in the middle — the undifferentiated model — and get repriced to the cost of electricity by the next weekend’s launch. Or you can own the one thing that behaves like gold: the scarce, graded, proprietary seam only you can dig. Your first-party data. A workflow whose output has a referee. The harness your customers actually live inside. Own that, and rent the commodity intelligence from whoever’s selling it cheapest this week, because there will always be someone selling it cheaper next week. Digital gold gets hoarded. Digital labor gets rented. Make sure you’re on the side of the ledger that appreciates.

What it means for you

Three moves, and they touch you whether or not you own a single AI name.

Route by task, not by reputation. Stop paying frontier prices for work a cheaper model finishes just as well. Opus 5 passed Fable and GPT-5.6 undercut both because most jobs never needed the top tier. Take one workflow running on your most expensive model and drop it a tier this week. If the output holds, the difference is yours to keep — and on this curve, it compounds.

Find your gold seam and guard it. The value that doesn’t deflate is the graded, proprietary kind. Coding has the compiler; your business has its P&L, its transaction history, its support logs. Name the one output only you can verify and only you own, and build there. Everything else in your stack is a commodity you should be renting, not defending.

Keep every model swappable. Loyalty is zero and prices get cut monthly, so portability is the strategy, not the vendor logo. The lab you’re on today will be undercut by the next launch, exactly as Fable was. Make sure you can move a workload — especially a coding workload — to a new model in a day, not a quarter, and every price war that follows pays you instead of trapping you.

The priciest model on earth just went quiet at 11 percent. When the canary stops singing, you don’t check the canary — you check the air. The frontier was supposed to be the rare, precious thing you paid up for. It turned out to be the bird in the cage. Own the seam that still pays, rent the rest on its way to free, and keep one eye on the air the whole industry is racing to stop measuring.

---

The email edition — THE NUMBER, the three moves, and the day’s five stories — is in your inbox. If the price of intelligence falling out from under you is forcing the question of which of your workloads are graded and which you’re paying a premium to guess at, that’s the conversation we run at Outsider Labs.