Anthropic – The issue of price

The cure for high prices is high prices.

• Weaker-than-expected demand for Anthropic’s flagship model is a sign that there is a limit to the price that the market will bear, especially when there are a larger number of pretty good alternatives available at much lower prices.
• Fable 5 is Anthropic’s latest, largest, best-performing model, but also by far its most expensive to date, and with everyone on a pay-as-you-go model, spending has flatlined.
• Data from Ramp AI Index indicates that enterprise spending on Fable 5 has flatlined, with spending shifting to Opus 5 (much cheaper) and other alternatives such as OpenAI and open source.
• Historically, corporate users have defaulted to the latest frontier model, but as spending on tokens has skyrocketed, those with their hands on budgets have questioned whether the latest and greatest is needed for everything.
• Instead, corporates are reserving their Fable 5 tokens for critical projects, with the day-to-day tasks being stepped down to cheaper alternatives.
• The net result was that Anthropic missed the most optimistic estimate that it would pass $80bn in annualised run-rate revenue (ARR), instead coming in at $65bn, already an impressive feat.
• This is clear evidence that the cure for high prices is high prices, and the softer cutting-edge demand is going to lead to prices for tokens starting to fall.
• This is something that the AI service providers are going to have problems with because the prices that they are paying for compute are at what I would consider to be supernormal levels.
• Hence, to make a return on the compute they have purchased, the likes of Google and Anthropic will need to increase prices for tokens substantially again, but I am not convinced that the market will bear it.
• The good news here is that the lower models consume far less compute than Fable 5 to produce a token, and so Anthropic may still be able to earn a good return as the traffic shifts to these far cheaper models away from the cutting edge.
• However, what it does do is reduce Anthropic’s competitive edge, as the superiority of Opus 5 over OpenAI or Open Source is much less clear and so it could begin to lose share, pushing it into a position where it has to cut its token prices.
• This is also likely to lead directly to softer demand for compute, meaning that prices normalise at around $20bn – $25bn / GW / year as opposed to the $30bn – $50bn where they currently sit.
• It is supernormal pricing that is enticing everyone to enter the datacentre market, and as supply increases in 2027, I expect pricing to begin to normalise.
• This is not necessarily bad news for Anthropic, as if both its costs decline faster than its prices, then its margins will remain stable or even expand.
• The only real losers here are those who are banking on token and datacentre pricing remaining where it is, as I suspect that they are in for a nasty surprise next year.