What Keeps AI Going: Open Models or Lower Prices?

Two leading providers cut prices at the same time

On September 22, Anthropic cut the price of Opus 5.5 by 20% and its caching price by 60%. OpenAI launched GPT-6 Sol and Luna at roughly half their predecessors' promotional prices, making the new prices permanent.

Meanwhile, OpenAI's compute infrastructure is expected to cost $856 billion over the next five years, with a cumulative funding gap of $278 billion.

AI is getting cheaper to use, but more expensive to build. Can price cuts and open models really sustain this spending race?

“Good enough” is changing the pricing logic

A Vercel report shows that open-weight models' share of usage rose from 7% last December to 56% in August. Over the same period, the average model price fell 23.2%. As more requests shifted to open-weight models, average model prices fell too.

The reason is straightforward: most tasks do not need a top-tier model; good enough will do. When a cheap model can do the job well, an expensive model struggles to maintain a high price. Users remain willing to pay more for tasks that are genuinely complex.

Usage shifted, but revenue did not follow to the same extent

The report shows that open-weight models handled 56% of AI requests but accounted for only 14% of model-call revenue. They had already won more than half of usage, but nowhere near half of revenue.

Consider Anthropic. Vercel's data show that, for every RMB100 users spent on AI calls, expensive models' revenue fell from RMB13.2 to RMB4.9 within one month, while cheaper models' revenue rose to RMB22.5. Anthropic's models together still captured RMB64. Users traded down, but most of the money stayed with Anthropic.

The pattern was similar within OpenAI. Luna handled more than eight times as many requests as Astra, yet generated only a quarter of Astra's model-call revenue. Cheaper models took on a large volume of routine tasks, while users remained willing to pay for high-end models when the tasks were difficult.

Prices can fall, but development cannot stop

This creates a dilemma. Providers need low-cost models to gain scale, while continuing to develop high-end models. Prices keep falling, but the next generation still requires substantial spending. That is where the funding gap emerges.

In August, Nvidia announced up to $105 billion in residual-value guarantees to backstop default risk for data centres leased by OpenAI. This is not a direct cash payment. Nvidia is using its own credit standing to reduce project risk and make outside capital more willing to participate.

It still comes down to funding

Compute efficiency will continue to improve, allowing the same amount of money to process more tasks. If efficiency gains outpace demand growth, funding pressure will ease.

For now, however, AI is being sold at ever-lower prices while still costing large sums to build. Companies must retain users with low prices while continuing to develop the next generation of models.

With prices falling on one side and spending continuing on the other, where does the money to bridge the gap come from? It has to come from outside.