Anthropic's price cut makes AI margin a routing decision

Fable 5.2 is a fifth cheaper, yet an 80-fold output price gap to Meta's Muse Ember shows margin settling with whoever splits work between model tiers.
Anthropic cut the price of its top public model by a fifth on 9 November. Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, against $10 and $50 for the version released on 1 September, and it writes up to 30% faster. The same day Meta released Muse Ember, a small model built for the execution steps of agent work, at $0.10 and $0.50.
Put the two price lists side by side and the gap on output is 80 times. That spread, more than the cut itself, is where margin in AI applications now gets decided. A product that sends every step of an agent's job to the frontier tier pays the top rate for work a cheap model handles. A product that plans with the large model and executes with the small one pays a fraction. Gross margin at the application layer is becoming a routing decision, and the companies that route well keep the savings the labs hand down.
Cost per task sharpens the point more than cost per token. Ember scores within half a point of Claude Haiku 5.5 on a common composite intelligence index at the same list price for standard prompts, yet it finishes a task on that index for about $0.14 against $0.41, because it spends roughly a third as many tokens. Efficiency now moves cost as much as price does. Founders building agents should measure cost per completed task, step by step, and treat the model behind each step as a line in the cost of goods that gets revisited with every release.
That discipline sits at the centre of what Fanaura invests in: application companies whose cost base falls with each model generation and whose value lives in workflow, data and distribution rather than in access to one model. Thin layers that resell a single model's output lose twice, once as labs ship the same workflows themselves and again as the price they mark up keeps falling.
The timing belongs to the argument. The cut lands days before Anthropic prices its listing, with revenue for the nine months to 30 September at $33.24 billion, about 13 times the same period a year earlier. A lab selling volume into an IPO book needs usage to grow faster than price falls, and the range set on 10 November already marks frontier AI at a lower multiple of revenue than private rounds paid.
The strongest objection is that cheaper tokens rarely mean smaller bills. Agents consume more tokens as they take on longer jobs, and a faster model invites more calls, so spend per customer can rise while unit prices fall. The signal that settles it is mix. Fable held about 11% of business spending on Anthropic's models in August. If that share climbs after the cut while application companies hold or widen gross margin, routing is where the value settles. If the share climbs and margins thin, the labs have kept the saving for themselves.
The labs will keep cutting the price of a token. The money stays with whoever decides which token to buy.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building agents that route work across model tiers to cut cost per task? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.