Up front: this is an opinion piece, not a valuation model. I don't have Anthropic's books, and I'm a customer, not an analyst. I also use their models daily and rate them — Fable 5 is genuinely excellent at the things it's good at.
But I've watched the last twelve months of this market from inside a working stack, paying real invoices, and the gap between how Anthropic is priced and how the business actually behaves has stopped looking like a rounding error. Here's the case.
1. The moat is a lead measured in months, not years
Anthropic's valuation prices in durable frontier dominance. But every cycle, a competitor matches or beats the top model. Sol shipping at roughly half the price and — in my hands, on my work, better — is the current proof. You cannot underwrite a decade of premium on a lead that resets every quarter and doesn't always reset in your favour.
2. It's a price-taker in a commoditising market
Tokens are converging on a spot commodity. And the buyer's switching cost is genuinely this small:
- const model = 'claude-fable-5';
+ const model = 'gpt-5.6-sol';If your integration is behind any kind of provider seam — and mine is, deliberately — changing vendors is a config edit and a redeploy. When that's the switching cost, pricing power evaporates. Anthropic's premium is the first thing to crack, and the July price gap says it's already cracking.
3. It doesn't own its own compute
Training and serving run on other companies' silicon. Its largest investors are simultaneously its landlords and its competitors. A company whose cost structure and capacity ceiling are set by rivals doesn't deserve an independent-champion multiple — it deserves a discount for the dependency.
4. The revenue multiple is detached from the margin profile
The valuation implies software-like margins. But inference is a heavy variable cost that scales with usage: more customers and more success mean proportionally more GPU spend. That's much closer to a utility than to a SaaS platform, and it's being priced like the latter.
The shape of it, with round illustrative numbers rather than their actual books — which I don't have:
SaaS margin shape inference margin shape
1x customers -> 1x COGS 1x customers -> 1x COGS
10x customers -> ~1x COGS 10x customers -> ~10x COGS
100x customers -> ~1x COGS 100x customers -> ~100x COGS
marginal cost -> ~0 marginal cost stays ~linear
scale = margin expansion scale = more GPU hoursSoftware earns its multiple because the hundredth customer costs almost nothing to serve. Inference doesn't work that way: serving is a per-token cost that grows roughly with usage. You can bend the curve with caching, batching and distillation — and every lab is furiously doing exactly that — but you can't flatten it to zero the way a database-backed SaaS does.
This is the part I feel directly as a customer. Every efficiency they ship — caching, batching, cheaper small models — is them handing margin back to me to stay competitive. Good for me. Not obviously good for a premium multiple.
5. No consumer moat
No default distribution, no operating system, no billion-device install base. Claude mostly reaches users through other people's surfaces. OpenAI has ChatGPT as a household brand and a habit; Google has everything and can staple a model to it. Anthropic rents attention, and rent goes up.
6. Enterprise revenue is concentrated and poachable
A meaningful chunk rides on a handful of large API customers and coding tools. Every one of those customers is technically sophisticated enough to dual-source — many are explicitly architected to. That's the whole point of a model router. Concentrated revenue with low lock-in doesn't warrant a defensive premium; it warrants a discount for churn risk.
7. Open weights are a floor rising underneath the whole market
Every capable open model narrows the gap for the "good enough" 80% of tasks. I've already moved my log-watchers, classifiers and taggers off frontier models entirely, because a small local model does them perfectly for nothing. The paid frontier gets squeezed into a shrinking premium tier while free-and-adequate eats the volume underneath it.
8. Cash burn against a brutal capex treadmill
Staying at the frontier means each training run costs more than the last, indefinitely, with no guarantee the next one wins. That's a business that must keep raising just to stand still. Whatever else that is, it isn't a mature-company risk profile, and it shouldn't carry a mature-company certainty premium.
9. "Safety" is a brand, not a margin
The safety positioning is real and it differentiates in press and in hiring. I'll even concede it shows up in the product — I route irreversible, expensive-if-wrong work to Fable partly for that reason. But buyers overwhelmingly optimise on capability and price, and a differentiator most customers won't pay extra for doesn't belong in the valuation as if they will.
10. It's priced for winning a market that may have no single winner
The multiple assumes oligopoly economics: a few players splitting fat profits forever. The likelier outcome is many capable models racing margins toward zero, with the surplus landing on customers and on whoever owns the compute. In that world Anthropic is a strong research lab with a good product — and nowhere near what the last round implied.
The catch: being overvalued isn't the same as being bad
Here's where I'll break from the usual version of this take. Every argument above is about price, not quality. Anthropic ships excellent models, the safety research is real work, and Claude Code changed how I build. None of that is in dispute, and none of it is incompatible with the multiple being wrong.
The other honest caveat: I've been wrong about this shape of thing before. "Commodity, no moat, margins to zero" was also the confident take on cloud infrastructure in 2010, and AWS turned out to have enormous durable pricing power via switching costs nobody predicted. If Anthropic builds real lock-in — agent infrastructure, proprietary tooling, data gravity, enterprise workflow depth — the thesis breaks. Claude Code is arguably that attempt, and it's a good one.
The verdict
Anthropic is valued like a monopoly-in-waiting but operates in a market with low switching costs, rented infrastructure, commodity output, and competitors shipping better and cheaper. Great research shop, priced like an inevitability it isn't.
The practical takeaway for anyone building on top of this: don't architect as though today's leader is permanent. Put a provider seam in your stack, route by task tier, and treat the model as a swappable part. The lab that's ahead will change. Your ability to switch on a config value shouldn't.
Strong opinions, held loosely, re-checked quarterly.
Want a provider-agnostic AI layer so a vendor's pricing decisions aren't your problem? That's most of what I get hired for — start a conversation, or see recent projects.