Free models now account for 80% of tracked AI token use, up from 20% 12 weeks ago
One router-based chart shows the flip, driven by cheaper models that cost pennies only when firms run them themselves, while a premium market for the best models may remain.
All-In with Chamath, Jason, Sacks & Friedberg · Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment FailsPaid AI tokens have lost share to free ones on one widely watched measure. Investor David Friedberg, a co-host of the All-In podcast, pointed to a viral Versal chart based on tokens tracked through software routers, showing use flipping from 80% paid and 20% free to 80% free and 20% paid in 12 weeks. He called that shift unprecedented in technology markets.
Fellow co-host and investor Chamath Palihapitiya twice said, "That's insane." Palihapitiya said that pattern fits Anthropic, and probably other paid models, where a few customers buy the costliest tokens, and said those buyers could face pressure from finance teams or shareholders to trade down.
A second chart placed Anthropic's Claude models high on price and performance, with cheaper models from other labs, including Muse, GLM, and Kimi, sitting lower on price. Palihapitiya said those were hosted prices, while running some of the models yourself can cost under 10 cents per million tokens, though Friedberg added you then pay for computing and bandwidth.
Friedberg predicted small models will handle 80% or 90% of corporate tasks. Co-host and investor David Sacks pushed back, calling Anthropic and OpenAI a stable duopoly and saying 10% to 30% of buyers will pay a big premium for the best model.
Questions people ask
What is a token, and what does token share measure?
A token is a small chunk of text, roughly three quarters of a word, that an AI reads or writes. Bills are measured in millions of tokens, so a million tokens is roughly 750,000 words. Share of tokens is a measure of usage flowing through software, not a count of jobs completed.
What is the difference between paid and free models?
Closed models run on a company's servers and you pay per use, like Anthropic's Claude or OpenAI's models. Open-weight models publish their inner settings so anyone can download and run them on their own computers, often for free. The trade is convenience and top performance versus cost and control.
Why would big customers switch to cheaper models?
Most business work, like drafting, summarizing, coding help, and customer support, may not need the smartest model. Once cheaper models get close enough in quality, finance teams may ask why they are paying much more. That can push companies to test cheaper paid models or free ones they host themselves.