Analysis · Baillie Gifford US Growth Trust plc
Fable 5.1 Reduces Certain Usage Charges: What Baillie Gifford US Growth and Scottish Mortgage Still Need to Prove
In the scenarios published by Anthropic, a typical charge would be 25% cheaper thanks to caching, without a general decrease in unit rates. For the shareholders of the holding companies, the question is who will keep the profit: the customer, Anthropic… or both.
Courtesy translation of the French original.
The 25% figure looks like a promotion. It actually describes something more subtle: Claude Fable 5.1 would produce a lower bill on a typical load because Anthropic has reduced the price of cache reads by 75%, to $0.25 per million tokens. The prices of new entry and exit tokens remain unchanged. This nuance is crucial for Baillie Gifford US Growth and Scottish Mortgage, two major shareholders of the private company.
Summary in 30 seconds
The fact. Anthropic estimates, based on four weeks of real-world usage in August, that Fable 5.1 is about 25% cheaper for a token-charged fee and up to 45% cheaper for highly intensive agents. The result depends on the context reuse rate; it's not a uniform decrease.
The holding read. Anthropic held a 6.6% stake in Baillie Gifford US Growth and a 2.8% stake in Scottish Mortgage as of July 31. On reported assets, this equates to a cumulative indicative exposure of approximately £530 million. Better inference economies can facilitate adoption, but they only create value if volume and margin grow sufficiently.
The unknown. Anthropic is not disclosing the impact on its revenue, gross margin, or cash burn related to this change.
Why this small tariff line matters to holding companies
Baillie Gifford US Growth increased its private position in Anthropic in the second quarter, making it its fifth holding at the end of July. Scottish Mortgage ranked it among its top ten holdings. As this is a privately held asset, the link between the announcement and the net asset value is neither daily nor automatic: valuations are periodic and based on comparable transactions and judgments.
The mechanism is nonetheless clear. Lower cost per task can increase the number of economically viable tasks, then support usage, revenue, and eventually valuation. Between each of these steps, one condition remains to be verified.
What exactly is decreasing?
An agent often reuses the same long context: instructions, documentation, history, or codebase. Instead of treating it as entirely new on each call, the provider can cache it and then reread it at a lower cost. Anthropic now charges $0.25 per million tokens for this rereading, while new input tokens remain at $10 and output at $50 per million.
The benefit is therefore concentrated on repetitive use cases. The more an application reuses its context, the greater the potential savings. A short, constantly updated dialogue will see little benefit. Anthropic says it bases its estimates on actual August activity with the default effort level; however, these are the company's calculations, not an independent audit.
The paradox of price: giving to expand
Passing on part of the profit to the customer might seem to reduce unit revenue. But if a previously too expensive task becomes profitable, volume can grow. For Anthropic, three scenarios are possible: the price reduction expands the market and improves margins; it expands the market but the profit goes primarily to customers; or it triggers a competitive response that compresses prices across the entire industry.
Why it's promising
Lower costs can make previously uneconomic uses viable, expand Anthropic's addressable market and support the value of both trusts' holdings if volume offsets the lower unit price.
Why it's not yet proven
No incremental volume, revenue, true inference cost or margin by model has been disclosed. A competitive response could transfer most of the benefit to customers rather than shareholders.