See When Your Chat's Prompt Cache Runs Out
September 30, 2026 · InfraScout Team
The chat header now shows a small countdown next to the cost badge. It tells you how long the conversation stays cheap to continue. Once that window closes, it tells you what your next message will cost to pick the conversation back up. Long investigations had this cost all along, and now you can see it before you pay it.
What the badge tracks
When a conversation runs on Claude, the provider caches what it has already read so the next request doesn't pay full price for it again. The conversation history stays cached for five minutes after each request. The system prompt and tool definitions stay cached for an hour. If you send a message while the history is still cached, Claude reads it back at the much cheaper cached rate. If you send it after the cache has expired, Claude first writes the whole history back into the cache. On a long conversation, that write is the most expensive part of the turn.
The badge counts down that five-minute window:
- While the cache is warm, it shows the minutes left, such as
4m. - In the last minute, it switches to a seconds countdown such as
0:42and takes on a warning tint. - After the cache expires, it reads
expired.
Every message you send restarts the window, so in an active back-and-forth the badge sits near 5m and you can ignore it.
What the tooltip tells you
Hover the badge for the details. While the cache is warm, the tooltip says how many minutes are left. Once it has expired, it says what your next message will re-cache, for example re-caches up to 45.1k tokens (≈ $0.17). Estimates under a cent show as < $0.01. Screen readers announce the same sentence, so the information doesn't depend on hovering.
The estimate grows with the conversation. A short chat costs fractions of a cent to re-cache. A conversation that has pulled in event logs, sign-in exports and tool output from a dozen hosts can carry a large history, and you pay that re-cache every time you come back after a break.
Using it in practice
Nothing goes wrong when the cache expires. The conversation carries on exactly where you left it, and the next message just costs one re-cache. The badge makes that visible, so you can decide with the numbers in front of you.
If you're in the middle of an investigation and about to step away, you know that coming back after five minutes adds one re-cache to the bill. If you have a follow-up question ready, ask it before you leave. If you're done with a line of inquiry and the context has grown large, starting a fresh conversation for the next topic is often cheaper than reviving the old one.
When the badge is hidden
The badge appears only when it has something accurate to show. It stays hidden:
- While a turn is running, because the window restarts as soon as the turn finishes.
- On ended or interrupted conversations.
- On OpenAI and Azure AI Foundry (OpenAI) connections, which don't have a cache window like this to show.
- For anyone other than the conversation's owner.
Why not just cache for an hour?
Claude also offers a one-hour cache, and caching the history that long sounds like the obvious fix. The catch is the price of each write. A five-minute cache write costs 1.25 times the normal input rate, and a one-hour write costs twice the normal rate. Every tool step in a turn writes new tokens to the cache, and a single InfraScout turn can run many tool steps. When we modeled a blanket one-hour history cache against how people actually use chat, it came out about a fifth more expensive overall. The extra cost of every write outweighed the occasional re-cache it saved.
So InfraScout keeps the history on the five-minute cache, keeps the system prompt and tools on the one-hour cache, and shows you the countdown.
More on conversations, costs and the chat header: Chat.
Questions or feedback? Reach us at info@infrascout.cloud.