Remember when OpenAI was the obvious default for any company doing serious AI work? That tidy assumption got a lot messier this year, and the latest data from corporate spend tracker Ramp shows the pendulum swinging once more — back toward the company that started it all.
Ramp, the corporate card and expense management company, tracks what tens of thousands of American businesses actually pay for AI. Its newest numbers, covered by TechCrunch on August 20, show Anthropic still leads OpenAI among those paying customers. But the more interesting signal is the direction: OpenAI is gaining on Anthropic with business users again, and the gap is narrowing.
Where the Numbers Actually Stand
Here’s the timeline from Ramp’s data, as reported by TechCrunch. OpenAI was once the runaway first choice for businesses. That changed in May, when Anthropic hit 41% market share among Ramp’s paying business users to OpenAI’s 39%. OpenAI never clawed the lead back. As of July, Anthropic sits at nearly 44% to OpenAI’s nearly 40%.
The data covers more than 70,000 US companies that spend billions through Ramp’s bill pay and corporate card products, so it’s not a perfect census of the whole enterprise market. It skews toward the tech-heavy crowd that a Silicon Valley card attracts. But as a directional read on how the AI pie is splitting, it’s one of the better public signals we’ve got — especially while both labs are privately held and guarding their revenue numbers ahead of their IPOs.
Why OpenAI Is Picking Up Steam Again
Ramp economist Ara Kharazian points to a specific catalyst: developer enthusiasm for OpenAI’s latest model. “GPT-5.6 Sol is really good, increasingly the choice for developers,” he said. That model word-of-mouth is turning into measurable spending growth in Q3 to date.
The flip side is Anthropic’s Fable tier, which Kharazian says “disappointed both in adoption and real-world application.” Fable is expensive by design — it’s built for a narrower set of high-stakes use cases, not for general chat. And Anthropic salespeople made things harder on themselves by telling Fable customers the company would retain their data for 30 days, a compliance-driven requirement that landed badly with enterprises already twitchy about their data.
It’s a useful reminder that in enterprise AI, price and data policy are product features. A model can be brilliant at benchmarks and still lose deals because it’s a budget line item that’s hard to justify, or because the data-retention terms scare the security team.
The Real Story: Nothing Is Sticky Anymore
Step back and the specific winners matter less than the pattern, and it’s a pattern I care about because I run this kind of tooling every day.
When Ramp first flagged Anthropic overtaking OpenAI back in the spring, Kharazian put it bluntly: “We have never seen a software industry as dynamic, where newcomers can disrupt market leaders in a matter of months, and where the pace of development overrides the typical forces of vendor stickiness.” Those were warnings that Anthropic shouldn’t count the win as permanent. Turns out he was right.
Businesses are willing to flop back and forth between labs as each one ships a new model. That is extraordinary. In traditional software, once a company standardizes on Oracle or Salesforce or AWS, ripping it out is a years-long project. In AI, the switching cost is low enough that a single strong release can swing real corporate spending in a quarter. That volatility should give both companies’ investors pause about how durable enterprise AI revenue really is.
What This Means If You’re Building on AI
If you’re an engineer or a founder wiring AI into your product, this is the most actionable part of the story, and it’s exactly the lesson from building your own AI model router with LiteLLM: do not marry a single lab.
A year ago you could argue the model was the moat and you should just pick the best one and build deep. The Ramp data argues the opposite — the best model today is a moving target that can change vendors within months. The durable work you do is the harness around the model: your prompts, your evaluation set, your tooling, your fallbacks. That’s the part that survives a model swap.
I wrote about this recently in my week using Codex more than Claude — the harness, not just the underlying model, is what made the difference day to day. If you build with a router and treat every provider as interchangeable, a market share shift like this one is an annoyance you shrug off. If you hard-code everything to one API, it’s an emergency.
The Trust Angle Nobody’s Talking About
There’s another layer here that connects to a broader thread I’ve been following: people are using AI more than ever while trusting it less than ever. That reconciliation is playing out in the enterprise spending data too.
Businesses aren’t dropping AI — far from it. The share of Ramp’s customer base paying for AI topped 50% in March and hit nearly 56% by July. The market is expanding. But the trust is conditional and cheap. Companies are paying for AI that demonstrably works for their specific workflow, and they’ll move the moment a rival proves out. That’s not brand loyalty; that’s a spot market with a lot of buyers and sellers fighting for each deal.
It also connects to the safety question. OpenAI voluntarily hit the brakes on its own development this month under its Preparedness Framework, and Claude Code flipped its auto mode to default and asked users to trust the classifier. Both are gambles that the market will reward restraint and reliability rather than raw speed. The Ramp data is early evidence that those bets are in play, for better or worse.
So Where Does That Leave Us?
The honest read of the Ramp numbers is that the enterprise AI leaderboard is genuinely up for grabs every few months. Anthropic overtook OpenAI this spring, OpenAI is clawing back by August, and nothing says a third player — or a strong open-weight model — can’t disrupt both before the IPO clock runs out.
For companies buying AI, the takeaway isn’t to pick a winner and bet the farm. It’s to keep your options open, watch price and data terms as closely as you watch capabilities, and make sure the harness you build isn’t locked to a vendor that could be a different leader a quarter from now.
The crown keeps changing heads. The smart money is on whoever builds the best harness around whatever model wins the next round.