Claude’s Premium Is Losing Its Customers

US companies are reportedly directing less of their AI spending toward Anthropic’s most capable model and more toward cheaper alternatives. Payments data collected by Ramp from 70,000 companies shows spending on the premium Fable 5 model plateauing at 11 percent of total outlay on Anthropic’s AI tools, according to the Financial Times reporting summarized by Futurism. Customers are reserving the expensive model for difficult work while routing ordinary tasks to older Anthropic products, cheaper rivals, or open-weight models.

This does not establish that capability has stopped improving or that Anthropic lacks customers. It establishes a procurement habit. Once several models can summarize a document, draft serviceable code, or classify a ticket, the buyer stops asking which system is smartest in general. The useful question becomes which task deserves the surcharge.

The commodity forecast

On one reading, model intelligence is moving toward low margins. Providers spend heavily to train frontier systems, but customers can split workloads and buy only the small portion of performance they need. GPT-5.6 is reportedly cheaper to use than Fable 5, while open-weight competitors give technically capable firms another bargaining chip. If quality gaps keep narrowing, the frontier becomes a showroom: admired, occasionally rented, and unable to carry the whole business.

The weakness in that forecast is that enterprises do not purchase tokens in a frictionless market. They test security, negotiate contracts, build evaluation systems, connect internal data, train employees, and rewrite software around a vendor’s interfaces. A cheaper model can still be expensive to adopt. Reliability, latency, rate limits, compliance work, and migration failures all belong on the invoice, even when they do not appear in the advertised price.

The platform forecast

Anthropic could therefore lose premium-model share while preserving substantial power. Proprietary coding tools, agent systems, cloud partnerships, data-governance controls, and long enterprise agreements can make Claude the place where work happens, regardless of which model handles each request. The castle need not own every tool if it controls the gate, the ledger, and the road out.

There is also a less dramatic possibility: companies may be learning to use AI efficiently. A portfolio in which cheap models perform routine work and premium models handle rare, consequential tasks could expand total consumption while reducing the flagship’s percentage. Anthropic’s annualized revenue reportedly reached $65 billion in July, below an $80 billion investor estimate but still evidence that a plateau in one product tier is not the same as collapse.

The next few quarters should separate a temporary price war from a durable transfer of power. Watch premium-model spending per customer, price cuts, inference margins, open-weight adoption, contract length, and the share of workloads that firms can move without rebuilding their systems. If switching rises while margins fall, intelligence is commoditizing. If model prices fall but customers remain inside the same vendor stack, the landlord has merely discounted one room.