Aug 2026 · The Watch
Nobody upgraded to the best model.
If you’ve felt vaguely guilty for not moving your team onto the newest, smartest model — stop. Credit-card billing data from 70,000 companies says almost nobody else did either. And the reason is exactly the one you suspected.
The receipts.
Ramp’s AI Index tracks actual corporate card spend across roughly 70,000 companies — not a survey, not vibes, but money that moved. For July 2026, the split across Anthropic models looked like this:
| Model | Share of spend | Note |
|---|---|---|
| Opus 4.8 (older) | 28.0% | the workhorse |
| Fable 5 | 8.0% | “cost has made it a less popular model” |
| Opus 5 (newest) | 3.5% | released July 24 |
Read that bottom row again. A brand-new frontier model, out for a month, capturing 3.5% of spend — while the previous generation holds eight times that. This is not a slow rollout. This is a market declining an upgrade.
Scruuge’s TLDR
The best model available and the model people actually pay for are two different things. Your peers already voted, with their credit cards, for good-enough-and-cheaper. You’re not behind for staying put — you’re behind if you never made the choice deliberately.
It’s explicitly about price.
The analysis doesn’t leave the cause to guesswork. Fable 5 — by most accounts the strongest model in the lineup — sits at 8% because, in the index’s own words, “Fable’s cost has made it a less popular model.” Not “users found it unhelpful.” Not “quality disappointed.” Cost.
And this is happening while the vendor is growing fast — Anthropic’s annualized revenue reportedly climbed from roughly $47bn in May 2026 to about $65bn by July, with some 6,000 enterprise customers spending $100,000 a year or more. Demand for AI is not the problem. Demand for the most expensive tier of it is a different question, and the answer is showing up in the billing data.
What this actually licenses you to do.
The quiet fear behind a lot of AI spend is reputational: if I put my team on the cheaper model and something goes wrong, that’s on me. So the frontier default survives as insurance nobody priced. This data is the counterweight. The market has run the experiment at scale, and the overwhelming majority landed on the older, cheaper, good-enough option.
That doesn’t mean “always buy cheap.” It means the burden of proof flipped. Paying frontier prices should be a decision you can defend for a specific task, not the setting you never touched. That’s exactly what routing is — and given that over half of enterprise AI use is routine documentation, most of your volume was never a frontier job in the first place.
The honest caveats
- Share of spend isn’t share of value. A model used for 3.5% of dollars might be doing the hardest 3.5% of the work. Low adoption is not a quality verdict.
- New models start slow for boring reasons too — migration effort, evaluation cycles, rate limits, contract timing. Cost is a stated driver here, not the only one.
- It’s one payment rail. Ramp sees corporate card spend, which skews toward startups and mid-market, and misses committed enterprise contracts entirely.
Make it a decision, not a default.
The calculator shows what your current tier is costing you and which one fits your workload — two minutes, no sign-up. The $999 Assessment does it across your whole setup, task by task, so the frontier spend that remains is the spend you can actually defend.
Model-share figures from the Ramp AI Index (corporate card billing across ~70,000 companies, July 2026), surfaced alongside Financial Times reporting of 23 August 2026 citing people with knowledge of the matter; revenue figures are from that reporting. Scruuge’s contribution is the read and the caveats.