The Ramp AI Index: Business AI Usage Up 50% Since July, Spending Down — and Anthropic Now Has 51% of Enterprise Token Spend

The Ramp AI Index tracks AI spending from over 70,000 US companies through their financial data — a ground-level view of enterprise AI adoption from the accounts payable side rather than vendor press releases. The latest data, compiled through late September, shows a pattern that’s been developing since the summer: usage climbing, costs declining.

Usage hit a record high in late September. AI spending per unit has been declining since July — a trend Ramp economist Ara Kharazian calls “unusually long.”

The Numbers

Usage: Up approximately 50% since July, hitting a record high at the end of September.

Spending: Down overall despite higher usage. The decline reflects a sustained reduction in cost per token, not a reduction in how much enterprises are actually using AI.

Market share (token spending):

The Anthropic majority is notable: 51% of enterprise token spending, tracked through actual invoiced spend from US companies, not self-reported usage or marketing surveys.

What’s Driving the Cost Decline

Two dynamics are compressing costs:

Top-tier model price cuts: Both Anthropic and OpenAI have made price reductions on flagship models over the course of 2026. The competitive pressure is two-directional — each cut from one provider triggers matching pressure on the other.

Efficient model tiers: “Standard” and “lite” model tiers have become viable for a substantial portion of enterprise AI workloads. Deployments that previously required frontier-model capabilities can now run on smaller, cheaper models. The net effect is that companies’ average cost per AI unit of work is declining even when they occasionally use more expensive flagship models.

The combination of price cuts on premium models and efficiency gains from model tier optimization is compressing the average cost per token across enterprise portfolios.

The Open-Source Gap

Open-source models at under 5% of enterprise token spending is a striking data point given the amount of attention open-source models receive in technical coverage. Ramp’s data is specifically enterprise API spend, which skews toward large AI customers — this isn’t a representative sample of all AI usage, and consumer applications and on-premise deployments don’t appear in the data.

But for enterprise software teams procuring AI through API contracts: the <5% open-source share suggests that enterprise buyers are predominantly using managed services from Anthropic and OpenAI rather than self-hosting open-source models, at least in the API spending category that Ramp tracks.

The Sustainability Question

The “unusually long” characterization of the spending decline matters for planning. A brief price-compression period followed by price recovery would be one market dynamic; a sustained multi-month trend suggests the pricing floor is structural, driven by competitive dynamics that both Anthropic and OpenAI are committed to rather than tactical campaigns that will reverse.

If the efficiency improvements (better models per dollar, smaller models handling more workloads) continue at the current rate, the enterprise AI cost curve may continue declining even as usage grows — a flywheel that makes AI adoption economically self-reinforcing rather than budget-constrained.

Content created with AI assistance and reviewed for accuracy.