Why Cheaper Models Won't Win OpenAI the Enterprise

OpenAI is trying to chip away at Anthropic’s lead with enterprises by fashioning itself as a cheaper, more flexible provider, Steph and I reported this weekend. But Anthropic's impressive prospectus for an initial public offering, filed Monday, shows OpenAI will need to do a lot more than that to really move the needle.

To be sure, OpenAI has done a great job getting companies like Rippling to switch to its cost-efficient models from Anthropic’s. In fact, compared to Google and Anthropic, OpenAI has released the least expensive models this year when it comes to completing tasks in evaluations run by research firm Artificial Analysis.

For instance, the most cost-efficient model OpenAI released in 2026, GPT-6 Luna with a low reasoning setting, was 45% less expensive than Google’s cheapest model, Gemma E4B, and 80% less than Anthropic’s Claude Haiku 5.5.

But as we noted in our piece this weekend, the aftermath of Anthropic’s release of a not-cheap but very capable model, Opus 5.5, in late September took some of the wind out of OpenAI’s sails, at least among customers of OpenRouter, a service businesses use to access different models from one place.

That’s a not-so-surprising signal that model capabilities still remain a dominant factor in customers’ eyes. And it adds more urgency to the all-hands-on-deck effort at OpenAI to launch its next major model safely.

While Anthropic’s prospectus only covered the first half of the year and didn’t include the third quarter, it showed off strong business metrics. (We already know its growth rate slowed considerably between the second and third quarters, though it was still excellent. And Anthropic is almost certainly at least 40% bigger than OpenAI in terms of revenue.)

For instance, Anthropic said that as of Sept. 30, its net revenue retention was 900%, meaning it’s generating nine times more revenue from existing customers, on average, compared to the same period a year ago. (That figure excludes fixed-price Claude Pro and Max subscriptions, which Anthropic says don’t capture changes in spending the way usage-based products, which account for the vast majority of its revenue, do). In addition, the number of enterprises that spent $100,000 or more on Anthropic products in the past 12 months rose more than 12 times to around 11,900.

Anthropic is looking to lock in its lead by getting customers to sign longer contracts. The company said in its filing that more enterprise customers are signing non-cancelable, annual or multiyear contracts with minimum spend commitments. Right now, Anthropic doesn’t have such contracts with many of its top customers.

Given Anthropic’s commitment to leasing $657 billion in computing capacity over the next seven to 10 years, it may want to lock in some long term contracts with its customers.

Fighting the Narrative

While OpenAI turns itself into a budget option for enterprises, Anthropic is fighting the narrative that it’s a budget buster.

But Anthropic lately has taken steps to show its models are also getting cheaper to use. It said last month that Opus 5.5 cost 40% less to complete a task than an earlier version of the model, Opus 5, and it said Monday that the model motivated “many” customers to switch to using that model rather than older Anthropic models.

And earlier this month, Anthropic said its newest Haiku model was 75% cheaper to run, compared to an earlier version of the same model. Also in its prospectus, Anthropic said its most capable model, Fable 5.1, was 25% cheaper to use than its earlier flagship, Fable 5.

For all the hype around the technology, Anthropic showed that the market has barely taken off. In its prospectus, Anthropic said its products account for less than 1% of the expenses of its largest Fortune 500 customers, and therefore their spending has a lot of room to grow. For other Fortune 500 clients that spend at least $100,000, but less than the top customers, Anthropic products account for less than 0.1% of their total expenses.

This represents an opportunity for OpenAI and other AI providers too. But first, these companies have to show that AI can take a share of a company's labor budget, not just replace legacy software or applications. While Anthropic itself has said its engineers are five times more productive now than they were just six months ago, due to Claude, the vast majority of companies won’t be able to match Anthropic’s efficiency for some time.

Here’s what else is going on…

Ex-GitHub CEO Launches Competitor

After GitHub experienced several months of prolonged outages earlier this year, many customers bemoaned the fact that there were so few viable alternatives, we reported this spring.

Now, former GitHub CEO Thomas Dohmke—who left the Microsoft-owned developer services firm last summer—has officially launched a competitor to GitHub.

Dohmke’s company, Entire, on Monday announced a new coding repository product where customers can upload and collaborate on code, similar to how people work together in a Google Doc. But unlike GitHub, Entire, which has raised $60 million from Felicis and other backers—says it was purpose-built for AI agents, including features that make it easier to track what AI coding agents are doing and explain why they took specific actions.

Whether Entire can take a bite out of GitHub’s pie remains to be seen, but Dohmke is betting that his former employer is ripe for disruption.

—Aaron Holmes