Welcome back to The Prompt,

OpenAI’s popular agent Codex and ChatGPT Work hit 20 million weekly active users last week. Part of it might have to do with cost. The ChatGPT-maker also announced that as its models become more efficient to train and run, it’s dropping pricing of its latest model GPT-5.6 Sol by 20% over the next three months. In late July, the company said it would make other models like GPT 5.6 Luna and Terra much cheaper too, with Luna becoming 80% less expensive.

Thibault Sottiaux, head of product and platform at OpenAI says frontier models serve as an important flywheel to keep pushing down costs for previous generations of models.

“When we push on the capability of these models, what we have the opportunity to do is to take that capability and then use it in order to make large improvements on smaller models and make those more cost-efficient,” he says.

That also includes making the company’s use of computing resources more efficient. OpenAI plans to spend $750 billion on data centers through 2030. “It was often pointed out as over-investing,” Thibault says. “We wish we had more compute. Everyone is compute strapped right now.”

Making its models more cost-efficient to run is crucial for OpenAI’s business, especially as it prepares for an IPO. The company’s revenue run rate is up 35 percent this quarter, according to a person familiar with the company. And enterprise revenue run rate (which is more than consumer ARR) is up 50% this quarter.

Now let’s get into the headlines.

In a 12-page essay,........

© Forbes