OpenAI Hits a Billion Users, Celebrates by Charging Almost Nothing

OpenAI announced this month that it has surpassed one billion active users, a population that would make ChatGPT the third-largest country on Earth, ahead of the United States and behind only India and China, neither of which lets you generate a Studio Ghibli version of your dog.

The company marked the milestone the way one does in the year 2026: by slashing prices. GPT-5.6 Luna’s API rates fell 80%, from $1 per million input tokens to $0.20, with output down from $6 to $1.20. Terra, the mid-tier model, got a 20% trim to $2 and $12. The flagship, Sol, held its price but gained a “Fast mode” — 2.5x the speed for 2x the price, for customers who wish to be disappointed sooner.

CFO Sarah Friar explained the logic with admirable circularity: “Better intelligence drives broader adoption. Broader adoption supports more investment.” Which is either a flywheel or a treadmill, and the distinction depends entirely on numbers OpenAI does not publish on purpose.

Some numbers it does publish, or at least leak in an orderly fashion: of roughly 900 million weekly users, about 50 million pay. That is a conversion rate of roughly five and a half percent, meaning the most transformative technology of the century converts at approximately the rate of a decent email newsletter. The 2025 financials — $13.07 billion in revenue against losses that outlets report as anywhere from $21 billion to $38.5 billion, a $17 billion discrepancy that nobody involved seems eager to reconcile — suggest a business model best described as “compute philanthropy with a tip jar.”

The price cuts are made possible, the company says, by real engineering: GPU kernel optimizations cut serving costs 20%, speculative decoding improved token efficiency by 15% or more. The machines are getting cheaper to run. The question is whether they’re getting cheaper to run faster than OpenAI is making them cheaper to buy, and the loss figures vote no.

But this is not really an OpenAI story. It’s a war story. DeepSeek launched V4-Flash at $0.14 per million input tokens — undercutting Luna’s brand-new 80%-off price on day one — then processed a reported eight trillion tokens in a single day, then announced it would raise prices sharply and add peak-hour surge pricing two weeks later. The frontier of artificial intelligence has arrived, economically speaking, at the business model of a parking garage.

The situation in China has grown so heated that Beijing has formally warned its AI industry against “involution” — the officially disfavored practice of competing so hard that everyone loses money forever, as previously perfected in solar panels and electric vehicles. When the state that built its economy on ruthless price competition asks you to ease up, you have achieved something genuinely special in the history of markets.

Anthropic, for its part, kept Claude’s prices exactly where they were, sitting out the discount war with the serene confidence of a restaurant that doesn’t do coupons.

The endgame is visible from here, and it is strange. Token prices are collapsing toward zero while the cost of producing them climbs toward the GDP of a G7 nation. A billion people use the product; a rounding error pays for it. Every lab is simultaneously the cheapest it has ever been and the most expensive it has ever been to operate, and the stated plan, universally, is that scale will sort it out.

One billion users. Twenty cents per million tokens. Multi-billion-dollar losses. Somewhere in that triangle is either the greatest business ever built or the most elaborate customer-acquisition bonfire in economic history, and the remarkable thing — the genuinely unprecedented thing — is that six years in, with a billion humans enrolled, it is still too early to say which.

Sources: