Ramp's October 2026 AI Index says business AI adoption is still rising, but growth in new customers and in dollars is slowing. In the update published October 14, 2026, Ramp's lead economist Ara Kharazian reports that among Ramp businesses in September, 43.8% paid Anthropic and 40.6% paid OpenAI, with overall AI adoption at 55.9%. His read: "it feels like we're at the top of the S-curve."
A line that spread alongside it says "Opus 5.5 drives 34% of business spend." We could not find that in the written report. This post separates what Ramp published from what is claimed elsewhere, and explains what the numbers mean for anyone choosing models or setting budgets. The primary source is Ramp Economics Lab's October update.
Claimed versus verified
| Statement | Where it comes from | Status |
|---|---|---|
| 43.8% of Ramp businesses paid Anthropic in September, up 0.25 points | Ramp October update | Published |
| 40.6% paid OpenAI, up 1.12 points | Ramp October update | Published |
| Overall AI adoption 55.9%, up 0.62 points | Ramp October update | Published |
| Top 1% of firms spent $9,568 per employee per month, down 3.1% from the July peak of $9,873 | Ramp October update | Published, with a volatility caveat |
| Two straight months of virtually no growth in token spend | Ramp October update | Published |
| "Opus 5.5 drives 34% of business spend" | A chart posted on X by Ramp's economist, relayed by coverage | Not in the written report we read; unverified by us |
| "OpenAI is capturing a larger share at Anthropic's expense" | Paraphrase in coverage | Partly: OpenAI gained faster, but Anthropic also rose |
A caution on the 34% figure: we could not find it in the written update, so we cannot tell whether it is a company share or a model share, or what population it covers. If you need to cite it, link the chart itself, not a summary.
What Ramp actually measured
Ramp's index uses spend data from its customers, American businesses paying through Ramp, so it measures paid adoption and dollars, not free usage, non-US firms or usage billed elsewhere. Key points from the October update:
- Adoption: Overall adoption is 55.9%, rising slightly faster than August's 0.40-point gain. Kharazian says he questions the value of the headline metric now, because more than half of businesses pay for AI and many pay for several models.
- Share movers: xAI is a small player but up 0.6 points on the popularity of Grok Bot. Mistral lags at under 1% adoption by US firms.
- Concentration: In September the top 1% spent $9,568 per employee per month. Per-employee spend rose 11.7% from the prior month but remains below the July high; Ramp notes the top-1% estimate is volatile and was revised up after adding new vendors.
- Early October: AI spend rose to a new peak, 1.5% above the previous peak in August, though growth remains well below earlier rates.
Why spend is flat while usage rises
The most interesting finding is the gap between tokens and dollars. Token volumes rose, but token spend did not: two straight months of virtually no growth. Kharazian's explanation is price declines at the frontier plus efficiency and quality improvements at every intelligence level, which push the blended price per million tokens down.
That fits the pattern we have been tracking. The price of a given capability keeps falling, and customers move to cheaper options. Our post on Opus 5 overtaking Fable 5 in enterprise spend showed how quickly spend shifts between models when a better one ships. Our look at startups saving money with open-weight models and Vercel AI Gateway's open-model token share show the other side: cheaper models absorbing volume.
For labs, flat revenue with rising usage is a warning about pricing power. For buyers, it is the reason to renegotiate rather than lock in.
Contracts: the customer has leverage
One reader question Kharazian addresses is whether OpenAI and Anthropic are pushing customers into contracts that lock in enterprise spend. Based on what he heard from businesses who signed or saw such deals, the customer is in the favorable negotiating position. Contracts are rarely signed assuming higher-than-normal spend, and where they are, labs offer discounts off sticker prices. He notes these could contribute to price cuts without driving more volume.
For a team buying AI, that supports a simple approach: ask for volume discounts, avoid commitments above what you already use, and keep the ability to switch. Ramp's separate post on model switching, mentioned in the update, argues switching is getting easier.
Where the Opus 5.5 claim fits
Even if the 34% number is right for a particular week or cohort, it does not mean Opus 5.5 holds a third of the market. Model-level spend shares depend on price: expensive models take a large share of dollars with a smaller share of tokens or customers. We have not seen the underlying table, so we cannot say how the chart defines share.
Opus 5.5 is a model we have covered from several angles, including its research-taste results on TasteVal, and demand for a flagship at premium prices is consistent with Anthropic's lead in paid adoption. But a dominant spend share would raise a risk question too, one we explored in Meta and Microsoft cutting internal Claude use.
Adoption, spend and tokens are three different things
The update is easy to misread because it mixes three measures, and each tells a different story.
Adoption counts businesses that paid a vendor at all. It is a headcount of customers. At 55.9% overall, it is approaching saturation among Ramp's base, which is why Kharazian calls the S-curve topped out for customer growth. A one-point gain for OpenAI can come from a few hundred new small accounts.
Spend counts dollars. It is dominated by the heaviest users: the top 1% of firms spend about $9,568 per employee a month. A few large customers changing their mix can move the dollar share of a model by several points without changing adoption at all. That is one reason a model-level spend share like the 34% claim can differ sharply from an adoption share.
Tokens count volume. Volume kept rising while dollars did not, which is the signature of falling prices rather than falling interest. If you only watched usage, you would conclude demand is booming; if you only watched revenue, you would conclude it is stalling. Both are true, and the difference is price.
When you read a statistic about AI "share", ask which of these three it is and which population it covers. Most disagreements between dashboards come from that mismatch.
Three scenarios for the next few months
- Price-led plateau continues. Frontier prices keep falling, usage rises, and dollar growth stays flat. Labs compete on discounts; buyers gain. This is the path Ramp's data points to.
- A new flagship resets spend. As happened when Opus 5 overtook Fable 5, a model that is clearly better can pull spend toward a higher price tier. Dollar growth resumes without new customers.
- Concentration risk bites. Because a small number of firms drive most spend, a few of them cutting back, as the Meta and Microsoft internal-use story hints might happen, could move headline numbers fast.
None of these is a forecast. They are the ways the same data could turn, and a reason to read the next update rather than rely on this one.
What this means for builders and buyers
- Budget on tokens, not seats. Spend per employee is dominated by a few heavy users. Track cost per task.
- Expect price drops. If the frontier price per million tokens keeps falling, annual commitments at today's rates are a risk.
- Multi-model is normal. More than half of businesses pay for AI and many for several vendors. Design for switching.
- Treat single-chart stats as leads. Check whether a share is of customers, dollars or tokens, and which population it covers.
- Remember the sample. Ramp customers skew toward US, card-spending businesses.
How to use this data in a budget review
If you own an AI budget, the update suggests a few concrete habits. First, split your own spend by vendor and by model, then compare it to the adoption picture: being one of the 43.8% paying Anthropic or the 40.6% paying OpenAI says little about how much of your bill each one deserves. Second, track cost per completed task rather than cost per token, because falling token prices can hide rising usage. Third, review commitments quarterly. Ramp's economist reports that labs concede discounts from sticker prices, so a contract signed months ago may no longer reflect the best price you could get today.
Finally, keep the sample in mind. Ramp sees card and bill-pay spend from its own customers, so firms that buy through enterprise agreements, cloud marketplaces or free tiers are under-represented. Use the index as a directional signal for the market, and your own invoices as the ground truth.
What to watch
- The next Ramp update, to see whether OpenAI's faster growth continues and whether token spend resumes growth.
- Whether a model-level table is published that confirms or corrects the 34% figure.
- Whether lab price cuts continue to offset usage growth.
- Any evidence that enterprise contracts become more restrictive, contrary to what Ramp heard.
Related reading
- Opus 5 overtook Fable 5 in enterprise spend: Ramp data
- Startups cut costs with open-weight models: Brex
- Vercel AI Gateway: open models at 78% of token volume
- TasteVal: Opus 5.5 research taste
- Meta and Microsoft cut internal Claude use
- What happened to the GPT-6 Astra hype
- Sources: Ramp Economics Lab, October 2026 AI Index update, Ramp AI Index
Figures are Ramp's as of its October 14, 2026 update and cover Ramp customers only.