A few months ago, I was running an AI implementation for a client, and within the first two weeks I noticed almost every new user was defaulting to the same model…whichever one was labeled “best.” Not because they’d tested it against a cheaper option. Not because their task actually needed that much horsepower. Just because it was there, and it felt like the safe choice.
We ended up building a whole approval process around that one observation. People had to justify why they needed the top-tier model before we’d unlock it for them. That single change cut the client’s AI spend by 40 to 50%.
That experience is exactly what came to mind when I read Ramp’s August 2026 AI Index. They titled it, bluntly, “Cracks in the AI Thesis.”
Real data, one important caveat
Ramp’s AI Index isn’t a survey, it’s built from actual, itemized transaction data across more than 70,000 U.S. businesses on Ramp’s corporate card and bill-pay platform. That’s a real strength. But it’s worth keeping in mind that Ramp’s original customer base skews toward tech and startup companies, so this dataset probably leans more technically savvy and more AI-adopted than the broader economy. Treat the trend as directionally right, not as a precise read on every industry.
With that caveat out of the way, the headline numbers are still striking. A year ago, roughly 35% of businesses in Ramp’s data had AI spend on the books. By March 2026, that number crossed 50%. Anthropic has taken the lead among providers, with 43.5% of businesses paying for its subscriptions or tokens as of July, just ahead of OpenAI at just under 40%. xAI is a distant third.
The best model isn’t the popular one
Here’s the part worth sitting with. Every provider now has a “frontier” model. The newest, most capable, most expensive thing they’ve shipped. And the data shows businesses aren’t reaching for it.
Anthropic’s flagship, Fable 5, launched at roughly $10 per million tokens, about double the going rate for OpenAI’s flagship, GPT-5.6 Sol. One month after launch, Fable 5 makes up only about 6% of the tokens businesses buy from Anthropic, and a little over 10% of the dollars. GPT-5.6 Sol, priced lower, is capturing closer to 25% of both tokens and spend at OpenAI and even that’s a minority of total usage there.
In other words: the majority of business spend, across both major providers, is going to models below the flagship tier.
You don’t use a Ferrari to haul gravel
That’s the analogy I keep coming back to. Some of these frontier models are genuinely remarkable — markedly better than anything else on the market for the hardest problems you can throw at them. But most day-to-day work isn’t the hardest problem you can throw at something. My own daily driver has always been Sonnet, and that’s what I recommend to clients too. When we run implementations, the training push is always toward the daily-driver model, not the flashiest one available.
You can use the Ferrari. You can use Fable 5. But if you haven’t tested whether it’s actually earning its keep, the odds you’re paying for capability you’ll never use go up fast.
What this means for how you budget AI
Treat AI as a vendor category, not a single line item. Every vendor prices differently, flat subscriptions, token-based usage, lower-cost-per-token platforms. You need visibility into spend and token volume by vendor and by model, not just a lump sum.
Require a cost-per-performance case before anyone gets access to the top-tier model. This doesn’t need to be a heavy process. In the implementation I mentioned, the bar was intentionally low, a short justification, maybe a quick conversation about the use case. The point wasn’t to block people. It was to add just enough friction that nobody upgrades on autopilot.
Watch spend per employee, but don’t read too much into it on its own. Ramp’s July numbers show the median business spending about $12 per employee, the top 10% at $650, and the top 1% at roughly $7,500. Some of that top-tier number is inflated by AI-wrapper companies passing token costs through to customers, high spend there doesn’t necessarily mean high ROI.
Key Takeaways
- Growth is shifting toward cheaper, mid-tier models — the frontier models are not where most of the money is going.
- A real price ceiling exists. Even the best model on the market can’t charge whatever it wants and expect businesses to follow.
- Budget AI as a multi-vendor category and track spend by vendor and by model.
- Before approving an upgrade to a top-tier model, require a specific cost-per-performance justification.
- High per-employee spend doesn’t automatically mean high ROI — some of it is pass-through cost.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: https://everydaycpe.com/courses/technical-finance/cracks-in-the-ai-thesis/
Leave a Reply