On 7 October 2026 Anthropic lowered the price of prompt cache reads on Claude Sonnet 5.5 from $0.20 to $0.10 per million tokens. That is now 0.05x the base input price instead of 0.1x. Cache write prices and every other price are unchanged.
Two days later, on 9 October, Anthropic added a beta of dynamic workflows for Claude Managed Agents, which lets an agent plan and run multi-phase jobs in the background, such as reviewing hundreds of documents, and combine the results.
Plain English: what is a cache read?
Most AI features send the same material with every request: instructions, a policy document, a product catalogue, a customer’s history. Caching lets the model provider store that material and charge a much lower rate when it is reused. The cheaper that reuse is, the less it costs to run a feature that leans on a lot of fixed context.
Who benefits
- Support and knowledge assistants that load the same help centre or policy text on every question.
- Document-heavy tools in healthcare, legal and finance where each query carries a large reference pack.
- Agents that repeat long instructions across many steps.
For a chatbot answering short, unique questions, the saving is small. The effect is largest where the same large block of context is reused thousands of times a day.
What to be careful about
Vendor prices have moved several times this autumn, in both directions. A saving this month is not a promise for next quarter, and only some designs are structured to take advantage of caching. Treat published token prices as an input to a running-cost estimate, not the estimate itself.
So what
If you are scoping an AI product, ask your supplier for a running-cost model per thousand users that states its caching assumptions, and a plan for switching models if prices or terms change. We build that into our AI product development work, and our AI-assisted development approach keeps model choice swappable. To cost an idea, start a project.