Anthropic cut the cost of cache reads in Claude Fable 5.1 by 75%, to $0.25 per million tokens, and left untouched the two numbers everyone quotes: $10 per million input tokens and $50 per million output, identical to the previous version. By the company's own measurement, four weeks of usage through August showed ordinary bills falling around 25%, and bills for agent-heavy work falling by as much as 45%. The deepest cut in the release landed on the one line of the rate card almost nobody mentions.
A chatbot takes a question and returns an answer. An agent works differently: it keeps going back to context it has already read.
The loop is the whole story. An agent reads instructions, a codebase, open files and the results of its own previous attempts, takes one step, then repeats. Every iteration pushes the accumulated context through the model again. A 40-step task sends roughly the same body of text 40 times. Input caching exists so that each of those passes is not billed at full rate.
In serious agent workloads, cache reads are where most of the tokens are. Fresh input and generated output are comparatively small. A price list showing only input and output tokens is therefore describing the minority of the cost of the exact workload the industry has spent the past year building agents for.
Holding published rates while cutting the real cost of running does two things at once.
The $10 and $50 stay the anchor for enterprise negotiations and model comparisons, and Claude's positioning stays where it was. Meanwhile customers running long agent loops, which is Anthropic's core competitive ground, watch their bills fall by a quarter to nearly half. Nothing changes in the figures procurement works with.
The claimed capabilities point the same way. On Terminal-Bench-Science, a test of research tasks run by an agent in a terminal, Anthropic reports 52.6%. The previous version scored 24.7%, and Claude Opus 5 scored 29.0%.
That is one strategy, not two: the model gets better at sustained terminal work, and sustained terminal work gets cheaper in the same release.
The assumption that lower prices mean lower spending has not held up in this cycle, and the mechanism behind that is old. When a unit gets cheaper, buyers use more of it, and total consumption can rise even as the per-unit price falls. William Stanley Jevons described the effect in coal consumption in 1865. It fits AI inference: average token prices fell roughly 67% year over year between the first quarters of 2025 and 2026, and inference costs dropped another 43% over ten weeks this summer. Industry-wide compute spending rose across the same period.
Agents sharpen the effect. A developer who could previously afford a 20-step run can now afford a 40-step one for the same money. Longer loops solve harder problems, so they get used more often, and the cheaper the loop, the longer people are willing to let it run.
Which is why the companies cutting token prices are signing hundreds of billions of dollars in compute commitments in the same breath. In the last week of August alone, Anthropic signed roughly $80 billion in agreements, including a $45 billion contract with Nscale. A company expecting its own price cut to shrink demand would not take on those obligations.
My read is that this discount was engineered to be invisible to exactly the people who would otherwise use it in a negotiation. Published rates are what a competitor undercuts and what a buyer screenshots. The cache-read line is what the bill actually runs on. Keeping one fixed while cutting the other lets Anthropic hold a premium position and compete on price at the same time, and the only people who notice are the ones already running the workload.
Notably absent from the framing is any figure for what share of agent tokens are cache reads. That is the single number that would let a customer verify the 25% and 45% claims rather than accept them. The savings are self-reported, measured by the vendor, over a four-week window the vendor chose, and "as much as 45%" is a ceiling presented where a typical case belongs.
The money freed by a price cut does not vanish. It moves to whoever controls the scarce resource. Companies building agentic software get a margin increase on work they are already doing, with nothing to renegotiate and no code to change. The compute layer gets the extra volume that cheaper loops create. That is why demand for accelerators keeps outrunning supply even as the price of the compute they produce falls.
Nvidia expects roughly 70% revenue growth in fiscal 2028, well above the 44% analysts forecast. The company's finance chief attributed that figure to constrained supply rather than thin order books.
To read an AI price cut correctly, find the rate-card line the volume actually passes through. Anthropic made agent work substantially cheaper without touching the numbers that reach headlines, and the clearest beneficiaries sit one layer down, selling capacity the cheaper loops will consume. That leaves Anthropic in the position it has chosen for itself: it has taken the cost out of the loop and committed $80 billion to the machines the longer loops will run on. The cut only pays for itself if the demand it induces is as large as the company is already behaving as though it will be.