Anthropic’s flagship Fable 5 model accounts for roughly 11% of what companies spend on Anthropic’s tools, according to Ramp payment data covering 70,000 companies reported by the Financial Times on August 23, 2026. Enterprises are sending the other 89% of that spend to cheaper models, including DeepSeek, which prices comparable work 10x to 50x below Fable 5.
Fable 5 costs $10 to $50 per million input and output tokens. DeepSeek’s equivalent models cost under $1 per million tokens. The gap has already moved prices: average prices paid for Anthropic’s models fell about 25% since mid-July 2026. Anthropic still holds 41% of enterprise AI budgets as of May 2026 and an annualized revenue run rate near $47 billion, with Claude reaching roughly 245 million monthly active users by mid-2026. Anthropic declined to comment to the Financial Times. The data lands weeks before an IPO expected to value the company above $2 trillion.
“Most people don’t need to operate at the frontier. [The period in which customers tended to opt for only the frontier models] was not a durable era.” — Miles Clements, Partner at Accel, which has invested approximately $1 billion in Anthropic, speaking to the Financial Times
What Model Routing Means for AI Model Pricing in Enterprise 2026 Budgets
Model routing is the practice of matching each task to the cheapest model that completes it correctly, reserving frontier models for the small share of work that fails on cheaper ones. Routing converts model choice from a procurement decision made once into a runtime decision made per request, and it is the mechanism transferring enterprise spend away from premium tiers.
The economics favour routing for three workload classes: high-volume classification, document extraction, and templated drafting. Each runs at near-identical accuracy on a sub-$1 model, so paying $50 per million tokens buys no measurable output difference. Lindy AI moved its usage fully to DeepSeek on cost grounds. Buyers assembling a routed stack can compare the current field in our guide to the best AI tools for business, where per-token cost now sits alongside capability in the scoring.
How Anthropic Is Defending Its Price Floor
Anthropic launched Claude Opus 5 between June and August 2026 at $5 per million input tokens and $25 per million output tokens, a mid-tier product priced to keep routed traffic inside Anthropic rather than lose it to DeepSeek. Fable 5 itself launched in June 2026, was withdrawn after White House intervention on national security grounds, then relaunched.
The defence targets the middle of the curve, not the bottom. Open-weight models from Alibaba’s Qwen line remove API cost entirely for teams willing to self-host, and Chinese models currently lead intelligence-per-dollar comparisons. Anthropic’s remaining differentiators are enterprise trust, safety documentation, and compliance tooling — the criteria that decide which vendor runs the agent workloads covered in our roundup of the best AI agents for business tasks.
For Context: Our Coverage of AI Model Pricing
- DeepSeek V4 Pro at $0.87 per million tokens — the model release that set the price floor this story measures.
- Claude Opus 5’s launch pricing — Anthropic’s mid-tier counter-move, in detail.
- Echo AI at one-third of Fable’s cost — a second challenger targeting the same spend.
- Open models trail closed models by 3% — the capability gap buyers are pricing.
- Qwen3.8-27B self-hosting — the zero-API-cost option below DeepSeek.
- Anthropic’s free Fable 5 window in July — the earlier discounting move that preceded the 25% price decline.
An 11% frontier share across 70,000 companies is not a verdict on Fable 5’s quality — it is a verdict on how much frontier capability a normal business workload actually consumes. Treat the 25% price decline as the buyer’s window it is: renegotiate on renewal, benchmark two cheaper models against your own evaluation set before assuming the premium tier is required, and route by measured accuracy rather than by brand. Keep the frontier model qualified for the tasks that fail everything else, and keep your prompts and evaluations portable so the next 50x price cut costs you a config change instead of a rebuild.
