The Small Claude Opus 5 Change That Cuts Real Costs

Anthropic just cut the minimum prompt size for caching in half. Nobody is putting that on a thumbnail, and that's exactly why it matters.

With Opus 4.5, a prompt needed at least 1,024 tokens before Claude would cache it. With Opus 5, that floor drops to 512 tokens. Prompt caching lets you reuse the same instructions and context without paying full input cost every single time you call the model.

That sounds like a rounding error. It isn't, once you think about who actually hits that limit.

Small Agents Were Paying a Tax

Before this change, a small agent or a lightweight automated workflow had a choice. Either stuff the system prompt with extra context it didn't need just to clear 1,024 tokens and qualify for caching, or skip caching entirely and eat the full input cost on every run.

Both options are bad. Padding a prompt to hit an arbitrary threshold means you're paying to generate tokens you don't want, just to save money elsewhere. Skipping caching means every single call, no matter how repetitive, costs full price.

Cutting the minimum to 512 tokens removes that trade-off for a huge slice of real-world use cases. Smaller agents, narrower tools, single-purpose automations, all of them can now cache without the padding tax.

A discount that only shows up after the ten-thousandth run is still a discount.

Where the Savings Actually Show Up

Saving five cents once is nothing. Nobody restructures a workflow to save a nickel.

Saving five cents ten thousand times is a different story. That's the entire logic behind this update. It's not built for the demo you run once to show your boss. It's built for the workflow that fires every time a customer submits a form, every time a document lands in a folder, every time a scheduled job kicks off at 3am with nobody watching.

That's where automated workflows live now. Not as one-off experiments, but as background processes running at volume, silently, over and over. A cost improvement that seems irrelevant at the scale of one call becomes the difference between a workflow that's profitable and one that quietly bleeds margin.

If you're building or maintaining anything that runs Claude on repeat, this is worth an actual look, not a skim. Check whether your system prompts and reusable context blocks now clear the 512-token line where they didn't before. If they do, you've got caching available to you that wasn't there last week, and at volume, that adds up to something worth having.