Estimated read time: 7 minutes
Neither OpenAI nor Anthropic publishes financials. Both are private, both are inching toward eventual IPOs, and both have every incentive to describe their enterprise business in the vaguest possible terms. So when someone with real spending data speaks up, it is worth listening. On Thursday, corporate card and expense company Ramp released numbers showing that OpenAI has started closing the gap with Anthropic among American businesses, and the shape of that data says something more interesting than who is winning.
The short version: business customers are switching AI vendors fast. Not over years. Over months. If you run a small business and you have been agonizing over which AI subscription to commit to, that agonizing is the actual mistake.
Table of Contents
What the Ramp data actually shows
Ramp tracks spending across more than 70,000 American businesses that run bills and corporate cards through its platform. That gives it a view into who is actually paying for what, as opposed to who is announcing partnerships.
The timeline goes like this. OpenAI was the runaway leader with both consumers and businesses for most of the AI era. Then in May, Anthropic passed it among Ramp’s paying business customers, hitting roughly 41 percent share against OpenAI’s 39 percent. OpenAI has not retaken the lead since. As of July, Anthropic sits near 44 percent to OpenAI’s roughly 40 percent.
That reads like a story about Anthropic pulling away. The Q3-to-date figures say otherwise. According to Ramp economist Ara Kharazian, OpenAI is currently growing faster than Anthropic within this segment. There is still a month left in the quarter, which in this industry is a geological era, so the trend could reverse again before September closes.
Kharazian attributed the shift to product cycles, writing on X that GPT-5.6 Sol is “increasingly the choice for developers,” while Anthropic’s higher-end Fable 5 tier underperformed on adoption relative to its price and its data retention requirements. Anthropic drew criticism earlier this month when it told Fable users their data would be retained for 30 days under regulatory requirements.
That framing is probably too tidy. Fable is a premium tier aimed at a narrower band of use cases than a general-purpose chatbot, so comparing its adoption curve to a mass-market model release is not apples to apples. But the directional point holds: a single model release moved measurable spending inside one quarter.
Why the lead keeps changing hands
Here is the part that should reframe how you think about your own AI spending.
In most software categories, market share moves slowly because switching is painful. Move your CRM and you migrate contact records, rebuild automations, retrain your team, and eat three weeks of lost productivity. That pain is why software companies can raise prices and why investors pay up for recurring revenue.
AI model subscriptions do not work that way, at least not yet. The switching cost for a small business using a chat interface is close to zero. You cancel one subscription, start another, paste your prompts into a different box. Even for developers working through APIs, the abstraction layers have gotten good enough that swapping a model provider is a config change rather than a rewrite.
That is why the leaderboard flips every few months. There is nothing holding customers in place except current model quality, and current model quality changes with every release.
For OpenAI and Anthropic investors, that is a problem. Enterprise software is supposed to be sticky, and stickiness is a large part of what those valuations assume. Data showing customers flopping back and forth on a quarterly basis is an argument that AI subscriptions currently behave more like a commodity than a platform.
For you, it is an advantage. You are the one with the leverage in a market where nobody has lock-in yet.
What this data does not tell you
Worth being clear about the limits before anyone builds a strategy on these percentages.
Ramp’s customer base skews toward technology companies. It is a popular corporate card among startups and Silicon Valley firms, which means the sample overrepresents exactly the businesses most likely to chase the newest model. A dental practice in Tulsa is not making quarterly AI vendor decisions based on benchmark scores.
It also excludes large enterprises that manage spend through American Express or legacy procurement systems rather than Ramp. Those are the customers signing multi-year committed contracts, and those contracts are stickier by design. If you could see the whole market, the volatility would likely look milder.
And Ramp shared only percentages, not dollars. Share of paying customers is not the same as share of revenue. A vendor could lose share while gaining revenue if its remaining customers spend more, which is plausible for whichever lab is selling more premium tiers in a given quarter.
None of that invalidates the trend. It just means the honest reading is “meaningful signal from a tech-leaning slice of the market,” not “definitive scoreboard.”
The number that matters more than market share
Buried in the same dataset is a figure that should matter more to a small business owner than any head-to-head comparison.
The share of Ramp’s business customers paying for AI at all crossed 50 percent in March. By July it was approaching 56 percent.
That is the story. Not which lab is ahead, but that paying for AI has become the default rather than the experiment. A year ago, “we’re evaluating AI tools” was a reasonable answer. In a market where more than half of comparable businesses have a line item for it, that answer starts to sound like a competitive disadvantage.
It also means the market is growing fast enough that both labs can be gaining revenue while trading share. The pie is expanding faster than either slice is shifting.
What small businesses should take from this
Five practical moves, given a market that reshuffles quarterly.
Do not sign annual contracts for AI tools. The standard advice is to take the annual discount because it saves 15 to 20 percent. In a category where the best option changes every quarter, paying a year in advance buys you a discount on the wrong tool. Stay monthly until the market settles. It has not settled.
Keep your prompts portable. Your accumulated prompts, templates, and workflows are the only real asset you build in this category, and they are the thing most likely to get trapped. Store them in a document you control, not exclusively inside a vendor’s saved-prompts feature. If you have built custom GPTs or Claude Projects, keep a plain-text copy of the instructions somewhere neutral.
Run both for a month before you decide. Twenty to fifty dollars for a month of parallel testing is cheap compared to a year on the wrong platform. Give both the same three real tasks from your actual work. Not benchmarks. Your work.
Watch the terms, not just the model. The Fable data retention flap is instructive. Model quality gets the headlines, but retention policy, training-data usage, and regional data handling are what will actually cause you problems if you handle client information. Read those sections before you upload anything sensitive. Our guide to auditing your SaaS stack for AI risk walks through the questions worth asking.
Re-evaluate quarterly, not annually. Put a 20-minute calendar block at the start of each quarter to check whether your AI subscription is still the right one. That is roughly the cadence at which the answer changes right now.
Frequently asked questions
Does this data mean OpenAI is better than Anthropic?
No. It measures which vendor more businesses are paying, in one slice of the market, in one month. Model quality varies by task, and both labs leapfrog each other regularly. Test on your own work rather than on someone else’s share chart.
Should I switch AI vendors because of this?
Not because of this. Switch if the tool you are paying for is failing at tasks you need done, and a competitor handles those tasks better in your own testing. Market share is a lagging indicator of other people’s decisions, not a recommendation for yours.
Is it worth paying for AI at all for a very small business?
More than half of comparable US businesses now do, which suggests the answer for most has become yes. The better question is how much. A single 20-to-25-dollar seat used seriously usually returns more than three seats used casually.
What about Google, Microsoft, and the open models?
They are real and this dataset does not fully capture them, partly because Google and Microsoft AI spending often rides inside existing cloud or Office bills rather than showing up as a discrete AI charge. Treat the OpenAI-versus-Anthropic framing as a two-horse view of a wider field. Our rundown of AI agents covers more of that field.
How fast does this market actually move?
Fast enough that the lead changed hands in May and the growth rate flipped again by August. Plan your commitments around quarters, not years.
Related Coverage
- SaaS Tools and the AI Disruption Audit: how to pressure-test the software you already pay for
- AI Agents Explained: The 2026 Rundown: what the agent layer actually does beyond chat
- Best AI Writing Tools for Small Business: the practical end of the same buying decision
Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what’s next.