Anthropic overtook its own flagship
Written by AI · Translated by AI · Read the Swedish original
Anthropic's most expensive model took 11.4 percent of the money and 6 percent of the work. AT&T cut its AI cost 56 percent by routing easy questions to open models.
Anthropic's most expensive model, Fable 5, accounted for 6 percent of the tokens that companies in payments firm Ramp's card data bought from Anthropic during July. The Financial Times reported the figure on 23 August. At the same time the model took 11.4 percent of the money. Fable 5 costs twice as much per token as Anthropic's own Opus 5, so that gap mainly reflects the price list. The number that measures usage is the six.
Anthropic has also overtaken itself. Opus 5 arrived on 24 July at 5 dollars per million input tokens against Fable 5's 10, and according to the Financial Times had passed Fable 5 in corporate spending within a month. By how much is not disclosed. The spending did not leave Anthropic. It sits further down the company's own catalogue.
It looks like a slow launch. Ramp measures what companies buy, not how they decide, but the pattern points to a shift away from the name of the newest model and towards the task it is meant to solve. upplyst.ai described the rule on 16 August in Stop lighting cigars with a flamethrower. What is different now is that someone has done the arithmetic.
AT&T showed days earlier what it looks like from the inside. A piece of software in the middle judges how hard each question is and sends the easier ones on to open models such as Llama, Gemma and Nemotron. The cost of certain advanced tasks, coding among them, fell by up to 56 percent. Quality fell 2 percent by AT&T's own measurement. Open models already take 40 percent of employee questions and the target is 60 to 70. The Information reported the figures and PYMNTS summarised them on 20 August.
The saving makes it worth someone building that middle layer. AT&T is large enough to build it itself. A company with twelve employees is not. The model choice then belongs to the vendor, inside a service the company already pays for, and the price does not reveal which model answered. According to Statistics Sweden, 35 percent of Swedish companies with at least ten employees used AI during 2025, and among companies with 10 to 49 employees the share was 30.8 percent. The question for them is not which model is best, but who makes the choice on their behalf.
Two percent worse quality for a cost that fell by up to 56 percent is a deal AT&T said yes to. The problem does not belong to the buyer. It belongs to whoever has just built their strongest model and then watched customers pick the cheaper one in the same catalogue.
Ask upplyst.ai
Why does it matter?
The flagship is no longer the default that everything else is measured against. It is something you escalate to. That moves the question from which model is best to who decides which question deserves the expensive one. It also moves the problem to the seller: if the strongest model is right for six percent of the work, it gets harder to say what the next, larger one is for.
What is the background?
Anthropic launched Fable 5 on 9 June 2026. On 12 June an export control decision from the US Department of Commerce forced the company to cut off access for every customer, after Amazon researchers showed how the model's guardrails could be bypassed. The controls were lifted on 30 June and the model was available globally again on 1 July, which is why July is the first month where usage can be measured. Ramp's AI index is built on card payment data from 70,000 companies.
What is uncertain?
Ramp's data is card payments from 70,000 mainly American companies over a single month, not the whole market, and it does not capture what is bought directly through cloud providers. Fable 5's first month was also disrupted by the June shutdown. AT&T's figures are the company's own, reported by The Information and not independently verified, and the 2 percent quality drop is measured against AT&T's own numbers rather than a general benchmark.
What does it mean in Sweden?
The routing gain is available to anyone running their own infrastructure. According to Statistics Sweden, 35 percent of Swedish companies with at least ten employees used AI during 2025, and among companies with 10 to 49 employees the share was 30.8 percent. For them AI is usually a feature inside a service they already pay for, where the vendor has made the model choice and the price does not show which model answered. The Swedish question is therefore not which model is best, but who makes the choice and whether it is possible to find out.