The Enterprise AI Pendulum Swings Back: Ramp Data Shows OpenAI Regaining Ground on Anthropic

For two years, the story of enterprise AI adoption has been written as a one-way ratchet: Anthropic climbs, OpenAI slides, and the “Claude is the enterprise model now” narrative hardens into consensus. New spending data from Ramp — the corporate card and expense platform whose dataset has become one of the most-watched proxies for business AI demand — suggests the pendulum is swinging back. OpenAI, which lost the lead among Ramp’s paying business users back in May, is once again growing faster than Anthropic among U.S. businesses in the third quarter of 2026.

What the data actually says

The numbers, published August 20 and attributed to Ramp economist Ara Kharazian, cover more than 70,000 American businesses that collectively spend billions of dollars through Ramp’s bill-pay and corporate-card products. The topline trajectory looks like this:

• May 2026: Anthropic hits 41% share of Ramp business AI spend, overtaking OpenAI at 39% — the first crossover, and a moment that generated no small amount of schadenfreude directed at the ChatGPT maker.
• July 2026: Anthropic extends its lead to roughly 43.5% (nearly 44%), while OpenAI claws back to nearly 40%. On the surface, that looks like Anthropic consolidating its win.
• Q3 to date: OpenAI is growing faster than Anthropic in this segment. A month remains in the quarter — which, as TechCrunch’s Julie Bort wryly noted, is “like 30 AI years” — so the trend could shift again before the books close.

There are real caveats. Ramp declined to share actual dollar figures, only percentages. The dataset skews toward the tech industry, because Ramp is a popular Silicon Valley corporate card. And it excludes large enterprises that run spend management through providers like American Express. This is not a measure of the total market. But as a directional indicator from 70,000+ real businesses paying real invoices, it’s one of the best public windows we have into a market where the two dominant players are both private and both barreling toward IPOs.

Why OpenAI is bouncing back

Kharazian’s diagnosis, posted on X, is blunt: “GPT-5.6 Sol is really good, increasingly the choice for developers.” OpenAI’s latest frontier model has clearly landed with the developer segment that drives a disproportionate share of API spend. Sol’s positioning — strong coding performance at aggressive pricing — plays directly to the audience Ramp’s dataset over-represents.

Anthropic’s side of the ledger is more complicated. “Fable 5, meanwhile, disappointed both in adoption and real-world application given price + data retention requirements imposed by regulators,” Kharazian wrote. Fable, Anthropic’s higher-end model tier, is expensive — but as Bort notes, that may be an oversimplification: Fable is built for a narrower set of high-stakes use cases than a general-purpose chatbot, so raw adoption numbers undersell it. The sharper self-inflicted wound was Anthropic’s disclosure that it must retain Fable users’ data for 30 days, which triggered genuine customer outrage in a market where data-retention terms increasingly decide procurement.

The lesson isn’t that Anthropic fumbled or that OpenAI is invincible. It’s that enterprise AI buyers are behaving like consumers: they flop back and forth with every model release, chasing whichever lab shipped the better frontier model this quarter.

The stickiness problem

That volatility is the most consequential finding for both companies’ investors. Ahead of their respective IPOs — Anthropic reportedly targeting a valuation as high as $2 trillion as early as October, OpenAI plotting its own path toward 2027 — the bull case for both rests on enterprise revenue being durable, compounding, and defensible.

Ramp’s data undermines the “durable” part. A lead built in May can erode by Q3. A model that disappoints on price or data terms gets swapped out mid-contract. If a 4.5-point share gap can open and partially close within a single quarter, what exactly are underwriters selling? Recurring revenue that behaves like discretionary spending is worth less than recurring revenue that behaves like infrastructure — and right now, AI subscriptions look closer to the former.

There’s a countervailing force, though, and it’s the genuinely bullish part of the story.

The market is still expanding underneath them

Even as the two labs trade share, the overall pie keeps growing. The percentage of Ramp-tracked businesses paying for any AI crossed 50% back in March 2026 and reached nearly 56% by July. Both companies can grow revenue while fighting over percentages, because the denominator — businesses paying for AI at all — is still climbing.

That expansion is why share-swing headlines, while dramatic, undersell the commercial reality: this is not a zero-sum fight for a fixed pool of enterprise budgets. It’s a race to capture a market that has crossed from “early adopter” to “majority behavior” in under two years. Median AI spend per employee among Ramp businesses was around $12 a month in July, with the top 1% of spenders reaching far higher — meaning there’s enormous headroom as usage deepens from experimentation to production workloads.

What to watch

Three signals will determine whether Q3’s swing becomes a durable reversal or just noise:

1. Whether Sol’s developer momentum holds. Developer mindshare converts to API spend with a lag; if Sol keeps winning workloads, OpenAI’s Q3 growth rate compounds.
2. Whether Anthropic answers on data retention and pricing. Fable’s 30-day retention requirement was a regulatory-driven constraint, but how Anthropic restructures tiers for privacy-sensitive buyers will matter more than benchmark scores.
3. The August and September monthly prints. If OpenAI’s growth advantage persists through two more monthly snapshots, the May crossover starts to look like an anomaly rather than a regime change.

None of this diminishes how remarkable Anthropic’s rise has been — from single-digit enterprise share in early 2025 to parity and then the lead in mid-2026. But the Ramp data is a useful corrective to linear extrapolation. In enterprise AI right now, nothing is settled, loyalty is thin, and the standings can flip on a single model release. For two companies preparing to ask public-market investors to price decades of compounding advantage, that’s an uncomfortable truth — and for everyone else, it’s a reminder that the most interesting phase of this competition is just beginning.