The discount landed on the half of the bill you already controlled

The pricing line from the Fable 5.1 launch drew nearly as much attention as the benchmark table. Per the AI News issue on September 2, list pricing held at 10 dollars per million input tokens and 50 per million output, while cache reads fell from a dollar to 25 cents, a 75 percent cut. Artificial Analysis, quoted in the same issue, set the rest of the ledger beside it: roughly 1.7 times the output tokens, about a dollar forty saved per task by the cache change, and a net cost per task around 20 percent higher than the model it replaced. Two days later the same newsletter has Astra at 10 and 50 for standard and 20 and 100 for the fast tier, using about 10 percent fewer output tokens than GPT-5.6 Sol at max effort while charging 2.5 times the token price, which the aggregator scores as 75 percent more expensive per task.

The applause for the cache cut is not misplaced. Agent runs resend the same instructions, repository context, and prior steps over and over, and paying a quarter of what you used to pay to reread your own scratchpad is real money.

Notice which half moved, though. The vault’s context notes are specific about what caching rewards: stable prefixes, durable instructions first, targeted context next, volatile task and log output last. That is the half you already controlled, and arranging it that way was already the advice. The half that grew is generated tokens, and those are not bought by the prompt. They are bought by effort settings, repeated sampling, reasoning scaffolds, and how many steps a loop may take before somebody stops it.

The same page has a rule for that half. Buy the extra reasoning only when it improves a verified outcome, not because a long transcript looks thorough.

So the saving arrived on the part of the bill that was already an engineering decision. The growth arrived on the part most people never set on purpose.