The cheaper line on a model’s rate card may be the one your budget spreadsheet barely tracks.

In an October 25 analysis, practitioner Max Woolf reviews several months of OpenAI and Anthropic pricing and argues that cache-read charges deserve more attention in agent workloads. His post is new commentary on earlier announcements, rather than an October 25 price cut. He also argues that competing inference providers constrain prices. That is an economic interpretation, not evidence that any particular rate will last.

A cache read lets a request reuse previously processed prompt context at a separate rate. The difference is visible in Anthropic’s September 22 Opus 5.5 announcement: standard input costs $4 per million tokens, output $20, and cache reads $0.20. Anthropic reports a 40% reduction in typical workload cost against Opus 5. That is the vendor’s comparison, with its own workload assumptions.

Consider a deliberately simple example using those rates. One million ordinary input tokens, ten million cache-read tokens and 100,000 output tokens cost $4, $2 and $2 respectively: $8 before cache writes or other charges. The largest token count contributes only a quarter of that subtotal. Counting every token at the ordinary input rate would describe a different bill.

What should a team carry into its next model comparison? Record ordinary input, cache reads, cache writes and output separately. Then connect the total to an accepted result, including retries and human correction. A cheaper cache rate cannot compensate for a workflow that repeatedly loses reusable context or produces work someone must redo.

Woolf’s broader forecast may prove right or wrong. The operational question is already testable: replay representative tasks, inspect the billed usage categories, and compare the cost of results that meet the same acceptance standard.