Anthropic's Best Model Has 3.5% of Spending. Its Previous One Has 28%.

Anthropic launched Opus 5 on July 24 — its most capable model to date. One month in, it accounts for 3.5% of enterprise AI spending. The previous generation, Opus 4.8, has 28%. That gap is not an anomaly.

The Number

The Ramp AI Index tracks AI spending across 70,000 companies using Ramp corporate cards — a proxy for real enterprise adoption rather than developer experimentation. For July 2026, the breakdown of Anthropic model spending looks like this:

Sonnet 4.6 sits at 8.3% and Fable 5 at 8.0%. Anthropic’s overall revenue trajectory remains strong — annualized revenue hit $65 billion in July 2026, up from $47 billion in May — and the company expects its Q3 to be profitable. The revenue growth story and the model adoption story are compatible: enterprises are spending more on AI overall while being selective about which model tier they use.

How the Math Works

The Opus 5 gap follows a well-documented pattern in enterprise software adoption: version releases don’t immediately cannibalize their predecessors. The mechanism is workflow inertia. Enterprises that have integrated Opus 4.8 into production pipelines — prompt engineering, fine-tuning, system prompt optimization, downstream tooling — have a meaningful switching cost even if the new model performs better on benchmarks. A 10% quality improvement rarely justifies the engineering time required to revalidate every downstream integration.

The cost dimension compounds the friction. Opus-tier models are priced at premium rates relative to Sonnet and Haiku equivalents. Enterprises that optimized for Opus 4.8 at a specific cost-per-token point face a re-evaluation of unit economics when Opus 5 arrives at a different price point. The Fable 5 pattern, sitting at 8% despite being described in some coverage as “less popular due to cost,” reflects the same dynamic at a different tier.

Anthropic maintains more than 6,000 customers spending $100,000 or more annually. The revenue base is durable even if model-tier distribution leans toward the previous generation.

What It Signals

The most direct reading of the Ramp AI Index data is that enterprise AI adoption is model-version-agnostic in a way that the AI industry’s benchmark-focused narrative often misses. Enterprises buy capability at a cost point, integrate it, and then resist upgrading unless the improvement is large enough to justify reintegration costs. This is how enterprise software has always worked — the innovation adoption curve for an enterprise SaaS product looks nothing like a consumer app’s viral growth curve.

The corollary is that Anthropic’s revenue growth ($47B to $65B annualized in two months) is being driven more by new customer acquisition and expanded use cases than by existing customers upgrading to the latest model. That’s a healthy growth dynamic, but it means the economic moat for any individual model version is shorter-lived than the benchmark rankings suggest.

For builders architecting agentic systems: the Opus 4.8 dominance figure is relevant to supplier selection. The model that runs in 28% of enterprise AI spending is the model with the most production data, the most refined integration patterns, and the most realistic cost-per-task benchmarks. Benchmarks tell you what a model can do; spending data tells you what enterprises have trusted to run at scale.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.