Enterprises are buying cheaper AI models: what July's spend data reveals
Anthropic reached $65bn annualized revenue in July while its strongest model drew just 8% of enterprise spend. Fresh billing data shows buyers routing volume to cheaper models.
Anthropic just reported the strongest revenue run in its history, and in the same breath conceded something uncomfortable: its most capable model is not the one enterprise customers are paying for. The tension between those two facts is the most useful signal in this month’s AI market data for anyone setting procurement strategy.
The July figures
A Financial Times report drawing on people with knowledge of the matter puts Anthropic’s annualized revenue at $65 billion for July, up from $47 billion in May. The company expects its third quarter to be profitable, on the same accounting basis it used to declare the second quarter profitable. It also told investors that 6,000 customers now spend $100,000 or more with it annually.
The competitive picture shifted in parallel. OpenAI’s annualized revenue has jumped 35 percent in the quarter to date and now exceeds $40 billion, with the July launch of GPT 5.6 credited with jolting the company’s performance after a sluggish start to the year.
Both labs are growing quickly, so the headline race looks settled either way. The decision-relevant detail sits one level down, in what enterprises actually chose to buy.
What corporate card data shows
The report introduced many readers to the Ramp AI Index, which estimates model adoption from the billing activity of roughly 70,000 companies that use Ramp corporate cards. Its breakdown of Anthropic-related spend for July 2026 reads like a rebuke of the assumption that the newest flagship wins by default:
| Model | Share of July spend |
|---|---|
| Opus 4.8 | 28.0% |
| Sonnet 4.6 | 8.3% |
| Fable 5 | 8.0% |
| Opus 4.6 | 6.9% |
| Sonnet 5 | 3.6% |
| Opus 5 | 3.5% |
| Opus 4.7 | 1.7% |
| Sonnet 4.5 | 1.3% |
| Haiku 4.5 | 1.0% |
| Opus 4.5 | 0.7% |
Two rows deserve attention. First, Opus 5, released on July 24, took 3.5 percent of monthly spend, which is unremarkable given roughly one week of availability. Second, and more telling, Fable 5, widely described as the company’s strongest model, attracted only 8 percent of spend, while a previous-generation Opus accounted for more than a quarter of it. Observers parsing the data concluded that Fable’s cost has made it an unpopular choice for routine workloads.
The pattern generalizes beyond one vendor. When a capable-enough cheaper alternative exists, enterprises route volume to it.
Capability sells evaluations. Economics decide deployments.
Most AI selection processes still start with a benchmark comparison, and flagship models win those comparisons. But evaluation-stage preferences rarely survive contact with production economics. A workload running millions of requests per month prices every marginal capability point in dollars, and once teams can measure that trade-off, they usually buy the cheaper model that clears the quality bar for the task at hand.
That is exactly what the Ramp data captures. The heaviest spend concentrates on models that are one generation back: proven, integrated, and priced for volume. The flagship’s share grows slowly because it earns specific, high-stakes tasks rather than default routing.
The measurement gap most enterprises still have
There is a second lesson in the source of this data itself. Ramp could build this index only because model charges flow through cards it can observe. Most organizations have no equivalent internal view. They know their total AI line item, but not how spend distributes across models, teams, or use cases, and therefore not whether they are overpaying for capability their workloads never exercise.
This is becoming a board-level reporting problem. As AI budgets grow, finance leaders increasingly ask the questions vendors ask of each other: what did we buy, what did it deliver, and would a cheaper option have delivered the same outcome. Companies that cannot answer will find the conversation drifting toward blunt cuts. Companies with spend-level telemetry can defend productive spending precisely because they can show where it goes.
Three practices close the gap:
- Tag AI spend by team, use case, and model so the distribution is visible monthly, not annually.
- Set quality floors per workload instead of defaulting to the premium model everywhere.
- Review routing quarterly against price changes, since this month’s data shows vendor pricing moves faster than annual contracts anticipate.
What this means for strategy
The July numbers suggest a market maturing in a specific direction: adoption is broadening through cheap, adequate models rather than concentrating on premium ones. For enterprises, that reframes three common positions.
Waiting for the next flagship to justify expansion is a weak plan when the volume economics sit one generation back. Standardizing on a single premium model is hard to defend when peers demonstrably route most spend elsewhere. Treating model choice as a one-time decision ignores how quickly the spend mix shifts; the same index showing Opus 4.8 dominant today showed a different picture before May.
None of this diminishes the strategic value of frontier capability. It means capability belongs where the task demands it, and everywhere else the efficient option should win. Making that distinction operational, with measurement to support it, is the difference between an AI budget that compounds and one that merely grows.
If you want an outside view of where your organization’s model spend is going and whether it matches the value you are getting, book a strategy consultation with our team.