The Logo Lead Is Not the Invoice

• Two ledgers, one week
• Customer count is not the bill
• What the 13 / 8 split is saying
• The quieter crack underneath
• Distinct from the capacity stories
• What to put on the scoreboard

§Two ledgers, one week

On September 16, 2026, Ramp lead economist Ara Kharazian posted a weekly cut of the firm’s enterprise AI spend data: OpenAI’s GPT-6 Astra took about 13% of enterprise AI spend; Anthropic’s Fable sat at about 8%. Same week, the September Ramp AI Index still had Anthropic ahead on customer count — 43.8% of U.S. businesses paid Anthropic for subscriptions or tokens in August, versus 39.8% for OpenAI.

Those are not the same race. One counts logos. The other counts the frontier invoice.

§Customer count is not the bill

The Index is blunt about adoption. Anthropic extended its lead in August. OpenAI’s business footprint grew slower. Technical sectors still lead. Overall new AI adoption is rising and decelerating at the same time.

None of that tells you who is winning the expensive seat. A firm can keep a Claude subscription on the expense sheet and still route the hard agent work to Astra when the job actually burns money. Penetration is how many companies have a vendor relationship. Spend share is which model gets the dollars when the work is frontier-priced.

If your board slide only shows “who more businesses pay,” you are scoring a phone book.

§What the 13 / 8 split is saying

Kharazian’s weekly note puts Astra ahead of Fable on enterprise AI spend as of this week. He also says OpenAI’s growth is coming from shifts off Sol and some Anthropic models, plus net-new usage. That is a different claim from “everyone switched labs.” It is a claim about where the frontier dollar went after the September model stack landed.

Remember the price sheet. Astra and Fable 5.1 both list at $10 per million input and $50 per million output. Headline rates matched. Cache economics did not — Anthropic cut Fable cache reads to $0.25 while OpenAI’s cached input sits near $1.00 on the public table. So a spend-share lead for Astra is not “they were cheaper on the brochure.” It is that buyers still put more of the enterprise AI check next to Astra this week.

Score the completed-task cost and the intervention rate. The sticker price alone is advertising.

§The quieter crack underneath

The September Index also shows why this ledger fight matters now. Median per-employee AI spend in the top 1% of firms fell 9.7% last month, from $7,976 to $7,205. Ramp’s effective price per million tokens is down about 41% to $0.68 from a 2026 peak of $1.15 in March. Frontier models (Opus, Fable, Sol and peers) drove about 45% of token share — up from early July, down from a 53% August peak — while cheaper standard models carry more of the volume.

Summer vacation is a real confounder. So is procurement pushing company-wide defaults toward standard models that are “good enough.” Either way, labs that need rising frontier spend to fund the next training run are watching the wrong chart if they only celebrate customer counts.

§Distinct from the capacity stories

This is not the Pro-seat freeze. OpenAI pausing new $200 ChatGPT Pro sign-ups on September 10 was a consumer SKU capacity dial. This is enterprise spend mix.

This is not the cache-hit bill. Anthropic’s Fable 5.1 cache-read cut changed the operator spreadsheet for long agent runs. This piece scores who collected the frontier dollar after that pricing week, not how the cache line is metered.

This is not Claude Code’s weekly allowance reset. The September 14 permanent +25% / promo-exit cut is a subscription ceiling story. Spend share is what hits the corporate card when the agent is on the API.

And it is not the AI-safety standards talk. OpenAI, Anthropic, and Google discussing a FINRA-style body is coordination theater for another day. The Ramp numbers are the operator thermometer.

§What to put on the scoreboard

Keep both columns. Ask what share of businesses pay each lab. Then ask what share of this week’s enterprise AI spend sits on Astra versus Fable. If those columns diverge — as they do this week — your risk is not “Anthropic is losing customers.” Your risk is mistaking a logo lead for an invoice lead while frontier volume softens and effective prices fall.

The scarce object is not another launch-day leaderboard. It is whether your memo still treats customer count as proof of frontier demand when the spend table already disagrees.