Anthropic is making agent workloads meaningfully cheaper without touching the list prices that anchor enterprise negotiations. With Fable 5.1, the company cut cache read costs by 75% , a change it says lowers a typical bill by about 25% and an agent‑heavy one by as much as 45% based on four weeks of usage data through August. Claude Fable 5.1 still lists at $10 per million input tokens and $50 per million output — exactly what its predecessor charged — which is why the largest price cut of the release sits in the line item almost nobody quotes.

Why Cache Reads Dominate Agent Work

A chatbot conversation sends a question and gets an answer. An agent works differently, and the difference is entirely about what gets re-read.

An agent runs a loop. It reads its instructions, the codebase, the files it has opened and the results of everything it has already tried, then takes one step and repeats. The accumulated context travels through the model again on every iteration. A task with forty steps sends substantially the same body of text forty times, and cached input is the mechanism that stops each pass from being charged at full freight.

The consequence is that cached reads are the dominant token by volume in serious agentic work, while fresh input and generated output stay comparatively thin. A price sheet quoting only input and output describes the smaller half of the bill for the workload the industry has spent a year building toward.

A Competitive Move Wearing Technical Clothes

Holding the headline rate while cutting the operative one accomplishes two things at once.

The published $10 and $50 anchor every enterprise negotiation and every side-by-side comparison written about the model. They stay intact, and so does the positioning that goes with them. Meanwhile the customers running long agent loops, which is where Anthropic competes hardest, watch their invoices fall by a quarter to nearly a half. No change appears in the numbers a procurement team benchmarks against.

The capability claim points the same way. Anthropic measured the model at 52.6 percent on Terminal-Bench-Science , a test of agentic research tasks run in a terminal, against 24.7 percent for the prior version and 29.0 percent for Claude Opus 5. Better performance on long-running terminal work, priced specifically to make long-running terminal work cheaper, is a single strategy expressed twice.

The list price protects the position. The cache price wins the workload.

Cheaper Tokens Produce More Compute, Not Less

The intuition that falling prices reduce spending has been wrong throughout this cycle, and the mechanism is well understood.

When a unit of something gets cheaper, buyers use more of it, and total consumption can rise even as unit prices collapse. William Stanley Jevons observed the pattern in coal consumption in 1865, and it has described AI inference precisely. Blended token prices fell roughly 67 percent year over year between the first quarters of 2025 and 2026, and inference costs dropped another 43 percent over ten weeks this summer . The aggregate compute bill for the industry went up across the same stretch.

Agent workloads intensify that effect rather than moderating it. A developer who could previously afford a twenty-step agent run can now afford a forty-step run for similar money, and will use the longer one because it completes harder tasks. The cheaper the loop gets, the longer people are willing to let it run.

That is the reason the same companies cutting token prices are simultaneously committing hundreds of billions of dollars to capacity. Anthropic signed roughly $80 billion of compute agreements in the last week of August alone, including $45 billion with Nscale . A firm expecting its own price cut to reduce demand does not do that.

The money released by a cut like this does not disappear. It moves to whoever sits on the constrained side of the trade.

Software companies building agent products get a direct margin improvement on workloads they were already running, without renegotiating anything or changing a line of code. The compute layer underneath collects the additional volume that cheaper loops generate, which is why accelerator demand has kept outrunning supply while the price of what the accelerators produce keeps falling. Nvidia guided to roughly 70 percent revenue growth for fiscal 2028 , well above the 44 percent analysts had modeled, with its finance chief attributing the figure to supply limits rather than any shortage of orders.

Reading an AI price cut properly means finding the line that actually carries the volume. Anthropic just made agent work meaningfully cheaper without changing a number that anyone puts in a headline, and the companies best placed by that decision are the ones selling the compute it will consume.