The Best Model Money Can't Sell - August 23, 2026
Watch on YouTube · subscribe for the daily brief
Anthropic makes the best model in the world. Its own customers spend 89% of their budget on something else. The premium AI model just stopped being a premium business — and it happened to the company that makes the best one.
Yesterday's brief led with Anthropic's newly public numbers passing OpenAI — a $65B run rate, the safety-first lab as the revenue leader. Sunday, the Financial Times published the other half of that story, and it's the one that reframes the entire six-week arc. Because it turns out Anthropic makes the best model in the world, and can barely get its own customers to pay for it.
Here's the finding, and it's remarkable. Per FT reporting drawn from Ramp payment data across 70,000 companies, spending on Fable 5 — Anthropic's flagship, most-powerful model — has not surpassed roughly 11% of what those companies spend on Anthropic's tools. Eighty-nine percent of the money goes to the cheaper models. And the FT's key point is that this breaks the historical pattern: for decades, enterprise buyers gravitated toward the most powerful option available. Not anymore. The best model can't command the budget, even from the customers who already chose Anthropic.
Sit with why this is the real ending to everything we've tracked. All summer the drama was about capability — models escaping, deceiving, doing original math, tripping "Critical" thresholds. The implicit assumption underneath every one of those stories was that frontier capability is the prize, the thing worth racing toward and paying for. This FT data quietly says: the market disagrees. Enterprises don't want the smartest model. They want the cheapest model that's good enough — and "good enough" now covers the overwhelming majority of real work. The mechanism is the one this run has named repeatedly: model routing. Companies send simple tasks to cheap models (DeepSeek, GPT-5.6 Luna, Opus 5's mid-tier) and reserve the expensive flagship for the rare hard case. The net effect, in the FT's framing: less traffic and less revenue flowing to premium offerings. Average prices paid for Anthropic's models have dropped nearly 25% since mid-July. The premium AI model, as a business, is deflating — and it's happening to the company that makes the best one, which is the most damning version of the story possible.
This is why the two days belong together, and why the pairing is the whole 2026 market in miniature. Anthropic is simultaneously the revenue leader AND watching its flagship's economics erode. Both true. The revenue is real (routing still runs through Anthropic's cheaper tiers), but the high-margin premium tier — the part that justified a rumored $2T valuation — is the part customers are abandoning. That's the tension every AI investor should be staring at ahead of the largest IPO on record: the company is huge and growing, and its most profitable product is the one people are actively fleeing. Anthropic clearly saw it coming — Opus 5, launched in July at half Fable 5's price, was the defensive move, and the "automated safety filters intervene 85% less often" pitch was partly about making the cheaper model usable for the work Fable was too cautious (and too expensive) for. The strategy is: cannibalize your own flagship before the competition does it for you.
Drew Breunig, flagged by Simon Willison today, is the voice catching up to what the routing data already showed — the analysis is finally converging on the reality that price sensitivity, not capability, is now the dominant force in enterprise AI procurement. And that reframes what matters for you, the builder, more cleanly than any breach story did. The skill that pays in this market isn't "use the best model." It's routing — knowing which task genuinely needs the flagship (few do) and which is fine on a model a tenth the price (most are). The teams still defaulting every call to the most powerful model are lighting money on fire; the ones with a real routing layer are the ones whose AI bills make sense. This is the Stripe-bought-OpenRouter thesis, now proven from the demand side: the value moved to the routing layer because the routing decision is where the money is saved.
The rest of Sunday was quiet, fittingly. Your one builder's-bench item is a good one: vLLM's "Exploring Speculative Decoding on AMD GPUs" — and it's thematically perfect for today, because speculative decoding is a cost/speed optimization (use a small fast model to draft, a big model to verify), and doing it on AMD is about breaking Nvidia/CUDA dependence. Both threads — cheaper inference, hardware independence — are the exact forces hollowing out the premium tier. If you run inference at scale, it's a genuinely useful technical read on squeezing more throughput per dollar, which is the only game that matters now.
The through-line, one sentence: on August 23 the FT revealed that Anthropic's flagship Fable 5 captures barely 11% of its own customers' spend while 89% flows to cheaper models — breaking the decades-old "enterprises buy the best" pattern — which means the real ending of this whole summer isn't a safety story or a capability story, it's an economic one: the premium AI model stopped being a premium business, and it happened to the company that makes the best one.
One thing to do today: if you take one habit from this entire six-week run, make it this — build (or audit) your model-routing layer, and stop defaulting to the flagship. Go look at your AI spend and ask, task by task: does this actually need the most powerful model, or is it running there out of habit? The FT just showed that 89% of even Anthropic's own customers' work runs fine on cheaper models. Your ratio is probably similar, and you're probably paying flagship prices for a lot of good-enough work. Route simple tasks to cheap models, reserve the expensive one for the genuinely hard cases, and measure the difference. The market has already voted with 70,000 companies' wallets. The engineers who internalize that "good enough, cheap, well-routed" beats "the best, everywhere" will run the AI budgets that actually survive contact with a CFO.
Sources covered
1 story · 3 sources (0 news, 1 company, 2 voices)