Anthropic Leads OpenAI in Ramp’s September 2026 AI Spend Index — and What the Token Price Collapse Says About Business AI

Ramp’s September 2026 AI Index — built from real corporate-card and bill-pay transactions, not surveys — shows adoption still widening while spend-per-seat softens and buyers route away from frontier models.

What Happened

Ramp’s September 2026 AI Index, published September 9 and built entirely from Ramp’s own corporate-card and bill-pay transaction data, puts Anthropic ahead of OpenAI in a straightforward business-adoption count: 43.8% of businesses in Ramp’s base transacted with Anthropic last month (up 0.34 percentage points month over month), versus 39.8% for OpenAI (up 0.09 points). Before any structural inference is drawn from that gap, the methodology must be understood: “share of businesses” here means the share of Ramp’s own customer base — a sample skewed toward startups, small and mid-sized companies, and tech-forward buyers — that ran a transaction with each vendor. It is not U.S.-wide market share, and emphatically not revenue share. A $20-per-seat startup subscription and an eight-figure enterprise contract count identically in a business-count metric. Anthropic leading on this measure does not make Anthropic the economy-wide winner.

The token-price signal is harder to dismiss on sample-bias grounds. The effective price of a million tokens across Ramp’s base has fallen to $0.68, roughly 41% below the March 2026 peak of $1.15. That move is large enough, and sustained enough across multiple months, to be directional rather than noise. Separately, frontier-model token share — Ramp groups Opus, Fable, and Sol in this category, and those are Ramp’s own labels — has slipped to approximately 45% of token volume from a peak near 53% in August, as standard models (GPT-5.6 Terra and Claude Sonnet, again Ramp’s labels) absorbed more of the load.

The top-1% per-employee spend figure requires the most caution: at $7,205 per employee per month (down 9.7% from July’s upward-revised $7,976), it is both small-sample and volatile — Ramp flags this explicitly, and the fact that July’s number was revised upward after publication is the tell that single-month readings here carry wide confidence intervals. Ramp also notes that August-period declines in its data tend to echo the November–December seasonal lull, which means a 9.7% one-month dip is a data point, not a confirmed trend. Open-source models appeared at 6.4% of AI-spending businesses and 3.6% of all businesses in the base.

The Structural Read

Three structural signals survive the caveats, and each maps to a different competitive dynamic in the AI stack.

First: the enterprise-versus-consumer model split is real, and the Ramp data makes it visible in hard spend. OpenAI dominates consumer mindshare — ChatGPT is the category name the way Google was — but within Ramp’s base of businesses expensing AI on a corporate card, Anthropic is ahead by four points. The buyer who runs a corporate card is not the same buyer who subscribes to a chatbot. Business procurement decisions weight reliability, API stability, safety properties, and enterprise support differently than a consumer choosing an app. Whether Anthropic’s lead in Ramp’s sample holds across the full economy — including large-enterprise IT budgets, which Ramp’s SMB-and-startup skew underrepresents — is genuinely unknown. But the direction of the signal is clear: in the segment Ramp can see, business buyers are tilting Anthropic.

Second: inference is deflating, not discounting. A 41% drop in effective token price since March is not a promotional campaign — it is a structural repricing of the commodity layer of the AI stack. For any company whose business model is selling raw model access, this is a margin compression signal, not a temporary headwind. The analogy is cloud compute in 2014–2018: prices fell every year, predictably, and the companies that survived did so by moving up the stack into software, workflows, and data — not by defending raw compute margin.

Third — and most strategically legible — the model-layer barbell is now visible in aggregate spend data, not just in strategy decks. Frontier-model token share has fallen eight points from its August peak not because frontier models got worse, but because buyers have learned to route: send the hard tasks to the expensive model, push volume to the cheaper standard model. This is exactly the architecture Harvey operates one layer up — reserve frontier for what genuinely requires it, commoditize everything else. When that routing logic shows up in aggregate transaction data across thousands of businesses, it means cost discipline has reached the median buyer, not just the sophisticated operator.

Three Implications