"Model Routing" on the Rise: OpenAI and Anthropic's Premium Pricing Models Under Pressure
0xBroomberg
Enterprise AI users are rerouting simple tasks away from expensive frontier models to cheaper alternatives — a practice called "model routing" that can deliver 5–10× cost efficiency gains, directly challenging the high-volume, premium-pricing logic underpinning OpenAI and Anthropic.
What is model routing, and why is it catching on now?
Model routing sends complex tasks to expensive frontier models and diverts simple tasks to cheaper alternatives. This means → companies stop paying top dollar for every query and start sorting by task difficulty.
Glean CEO Arvind Jain estimates roughly 95% of enterprise AI usage still runs on the most expensive frontier models — even for tasks that cheaper options handle just fine.
Cognition CEO Scott Wu says routing routine work to "good enough" models can yield 5–10× cost efficiency gains. In plain terms = the same job gets done at one-fifth or one-tenth of the old price.
How alarming are the compute bills?
Cisco Chief Product Officer Jeetu Patel laid out a back-of-envelope calculation: $200 in token usage per employee per week → roughly $10,000 per head per year → about $900 million annually for a 90,000-person company.
Cisco's own AI spending has already overshot its budget. 30,000 engineers now rely heavily on AI for product development; the company has been forced to reallocate resources and prioritize token consumption.
This reflects a cost curve so steep it "caught even the largest tech companies off guard." CNBC reports that CFOs and boards are now tightening controls on inefficient AI spending.
What does Cognition's "guarantee" mechanism reveal?
Cognition introduced an "AI productivity guarantee": if its coding agent Devin delivers engineering value below what the customer pays, the company will subsidize usage credits — up to $10 million.
Scott Wu framed this as a response to the industry's lingering ROI debate: "You can burn billions of tokens and accomplish nothing. Companies should chase output, not volume."
This means → the seller is starting to pay for outcomes rather than consumption. Pricing power is tilting toward the buyer.
Where is the business logic of OpenAI and Anthropic vulnerable?
Once enterprises shift high-frequency, low-difficulty tasks en masse to open-source or cheaper Chinese alternatives, frontier labs are left with revenue only from complex tasks. In plain terms = the most lucrative "volume" business gets siphoned off, leaving only "boutique" orders.
Both companies' business models — and IPO expectations — rest on the assumption of massive demand plus premium pricing. This means → if the volume leg shrinks, the valuation foundation itself needs re-examination.
Patel believes this will not fundamentally upend frontier labs — cutting-edge technology still commands value — but the pricing model will have to shift: labs must improve usage efficiency rather than simply raise prices.
What is the real unresolved question?
Frontier models' premium position on the most complex tasks is hard to dislodge in the near term.
The core suspense for the market: how large a share do "the other tasks" — the ones that do not need the strongest model — actually represent?
The answer to that ratio will largely determine the valuation ceiling for leading AI companies. This reflects a deeper truth: the fight over AI pricing power is, at bottom, a bet on how task complexity is distributed.
Content is for reference only, not financial advice.