AI token spend grew only 1.5% across August and September

AI token use kept rising while business spend on those tokens grew only 1.5% across August and September. That is the core signal in Ramp data covered by TechCrunch on November 6, 2026. In July, token spend had jumped 42% month over month. Two months later, usage and revenue for model companies were no longer moving together.

For builders and buyers, that split is the story. More tokens do not automatically mean more dollars for OpenAI or Anthropic. Price cuts, enterprise discounts, routers to cheaper models, and hard budgets are turning a seller's boom into a buyer's market. Unite.AI's write-up of the same Ramp AI Index numbers makes the divergence even sharper for October: spend re-accelerated, but volume raced ahead of the bill.

What Ramp actually measured

Ramp's corporate-card and bill-pay platform tracks more than 85,000 U.S. businesses and billions of dollars in corporate spend. Paid AI adoption is read from observed transactions, not surveys. Ramp notes that this method likely understates activity, because free tools leave no payment trail and employees may use personal accounts.

The token metrics sit on a narrower slice: customers using Ramp Token Spend Management with connected provider usage and billing data. That subset is not a perfect mirror of Ramp's total AI spend volume, though Ramp says the directional story matches. Unite.AI repeats those dataset boundaries clearly, which matters when investors quote the Index as if it were the whole market.

Ramp lead economist Ara Kharazian told TechCrunch the slowdown is not a demand problem. Volumes are up. The issue is how cheap and how disciplined that demand has become.

The August-September stall, then an October bounce

Across August and September 2026, token spend grew only 1.5%. July had grown 42% from the prior month. That contrast is the headline number TechCrunch leads with, and it is the number that should change planning assumptions for anyone modeling frontier-lab revenue as a straight function of usage.

October did not stay flat. From the week ending about September 23/25 to the week ending October 23, token spend rose 8.9%. Unite.AI adds that token volume climbed 37.1% in the same window. Spend returned to growth. Consumption still outran the check.

That pattern is what Notion co-founder Akshay Kothari captured in a short internal-data post TechCrunch cites: tokens go brrr, costs go ahhh. More work completed. Less money extracted per unit of work.

Prices fell hard enough to break the old model

Between those late-September and late-October weeks, the blended price per million tokens across OpenAI and Anthropic fell from $0.76 to $0.56. Unite.AI lists that as a 26.8% decline. OpenAI's side dropped 42.4%. Anthropic's dropped 15.7%.

Those cuts sit on top of explicit product moves. In September, OpenAI cut API prices roughly in half for GPT-6 Sol and GPT-6 Luna. Anthropic answered with significant reductions on Fable 5.1 and Opus 5.5. Anthropic's own filing said price cuts slashed revenue growth by about $600 million. Enterprise discounts meant to lock in spend are also landing as revenue headwinds when customers take the discount and still route aggressively.

Kharazian frames the result as a buyer's market. The old Big Tech playbook of subsidizing usage to lock customers in works poorly when Meta, Nvidia, Chinese open models, U.S. open-weight labs, and tools like Jev keep offering cheaper paths. For many workloads, he notes, you only need a simple yes or no. Route that to a lightweight chooser, and already-cheap tokens get cheaper still.

Concentration: the top 1% still sets the weather

Spending is concentrated enough that a small cohort can swing the whole curve. Ramp's October 14 analysis, as summarized by Unite.AI, said 1% of firms drove about 80% of spending on OpenAI and Anthropic, and identified two consecutive months of virtually flat token expenditure.

In September, median monthly AI spend per employee was about $9,568 inside the top 1% of firms paying for AI, versus about $813 in the top 10% and about $15.70 for the median firm. TechCrunch's interview numbers sit in the same neighborhood: the top 1% had been hitting near $10,000 per employee at summer peaks and came in just below that in August and September; median firms spend roughly $12 to $15 per employee per month; the top 10% spend roughly $600 to $1,000.

Kharazian told TechCrunch that a stutter in the top 1% is "sufficient to slow down revenue entirely, despite the fact that we're seeing rising volumes." Growth among median and top-decile firms is not enough to close the gap. He also doubts that a doctor's office or a car dealership will ever need to approach the $10,000-per-employee intensity that software and finance coding-agent buyers hit.

Unite.AI adds a useful caveat: the top 1% is a small, volatile group, and Ramp revised estimates upward after adding new AI vendors. Treat the exact peak as noisy. Do not treat the concentration story as optional.

Adoption is still spreading even as spend flattens

Paid AI adoption reached 55.9% of Ramp businesses in September. Anthropic was paid by 43.8% of those businesses; OpenAI by 40.6%. Share of token spend also shifted: Anthropic moved from 51.3% in the week ending September 25 to 60.5% in the week ending October 23, while OpenAI moved from 44.2% to 35.1%.

Share gains inside a pool shaped by falling prices do not fix overall spend growth. A provider can win relative share and still miss absolute revenue trajectories that assume July-style compounding.

Why sophisticated buyers spend less per token of work

Kharazian's mechanism list is practical, not theoretical:

• Defaults that keep expensive models off long-running tasks
• Employee budgets and spend caps
• Routers that push work from standard or light models to even cheaper light models
• Enterprise discounts that lower unit revenue even when seats stay
• Open and cheap alternatives from Meta, Nvidia, Chinese labs, and tools such as Jev

As companies get better at AI, they get better at controlling it. That is good for CFOs. It is awkward for model companies whose valuation stories still lean on hypergrowth spend.

Market context without making it the lead

The same TechCrunch piece notes investor mood around Anthropic's IPO talk and OpenAI's fundraising. Bloomberg had reported Anthropic's potential valuation talk down toward about $1.2 to $1.3 trillion from earlier $2 trillion chatter, with a high junk rating expected from credit agencies. OpenAI lowered a proposed valuation to $1.25 trillion from $1.4 trillion and cut fundraising goals by about $10 billion. Those moves are context. The Ramp Index is the evidence that usage and spend are decoupling for real buyers on real cards.

What this means for builders

If you ship products on top of OpenAI or Anthropic APIs, treat unit cost as a moving target, not a fixed line item. A 26.8% blended price drop in a month of calendar weeks can rewrite your margin model overnight. Design for multi-provider routing from day one. Prefer workloads that stay valuable when the model behind them gets cheaper, not ones that only pencil when a single frontier model stays expensive.

If you sell AI features inside software, do not assume your customers' token bills will rise in lockstep with their usage. They will fight that bill. Offer usage visibility, budget controls, and default-to-cheap paths, or your buyers will build those controls around you.

If you are evaluating open or mid-tier models for production, the Ramp data is a green light to keep A/B tests alive. The market already rewards thrift. Teams that route simple classification, extraction, and yes/no gates away from flagship models are doing what Ramp says advanced buyers already do.

What this means for buyers

You are in a buyer's market if you have volume and discipline. Ask for enterprise discounts. Instrument token spend per team. Set defaults that block expensive models on long jobs unless someone opts in. Measure cost per successful task, not tokens for their own sake.

Watch concentration risk inside your own org. If one team or one use case drives most of the bill, a single pricing change or model swap can swing the P&L. Spread workloads across providers where quality allows. Keep a cheap path for high-volume, low-stakes calls.

Do not confuse adoption percentage with spend intensity. 55.9% paid adoption across Ramp businesses is broad. Median spend near $16 per employee per month is not the same economy as the top 1% near $9,500 to $10,000. Most firms will never need the top cohort's intensity. Plan procurement for the intensity you actually have.

The djcroman take

Labelled opinion, separate from the Ramp facts: the interesting risk is not that AI demand collapses. The interesting risk is that demand stays strong while revenue per token keeps falling faster than volume rises. That is a healthy outcome for customers and a hard outcome for any valuation that needs July's 42% MoM spend growth to look like a trend rather than a spike.

Coding agents created the commercially obvious spend wave. Personal agents may not repeat it at the same intensity. Kharazian is skeptical they are as economically productive. Until another use case broadens high-intensity spend beyond today's concentrated top buyers, expect more of what August and September already showed: tokens up, bills soft, and CFOs learning the controls that investors hoped would never matter.

The bottom line

Businesses used more AI tokens while spending on those tokens grew only 1.5% across August and September 2026, after 42% month-over-month growth in July. October spend rose 8.9% while volume rose 37.1%. Blended OpenAI-plus-Anthropic prices fell from $0.76 to $0.56 per million tokens. Roughly 1% of firms still drive about 80% of that spend. Paid adoption is broad at 55.9%, but intensity is not. For builders and buyers, the takeaway is simple: design for falling unit prices, route aggressively, and stop treating token volume as a proxy for model-company revenue.

Sources: TechCrunch (Rebecca Bellan, Nov 6, 2026), Unite.AI, Ramp AI Index / Ara Kharazian comments as reported in those pieces, Akshay Kothari Notion quote via TechCrunch.