Incident: A timestamp that switched the cache off

Module 2 · Incident file · About 30 minutes

What happened

On 7 July 2026, Max Musing, founder and chief executive of Basedash, an AI business-intelligence product, published "How a single timestamp broke our prompt cache" on the company blog. Its chat and dashboard agents send long prompts: tool definitions, system instructions, then the conversation. The provider's cache should have served most of each request from the last one.

It never did. A line in the system prompt gave the model the current date and time, to the second. "Because the timestamp lived in the system prompt, the [tools, system] prefix was different on every request." So "we wrote a fresh cache entry each turn and never once read from one."

On their provider a cache write costs more than ordinary input: Anthropic's prompt-caching page lists 5-minute cache writes at 1.25 times the base input price and cache reads at 0.1 times. Musing's verdict: "We were paying the cache-write premium and getting none of the read discount, which is worse than not caching at all." The post does not say how long the line was live or give a total cost, so neither is stated here.

Once they saw it, they went looking for "everything else that was in the wrong place". The timestamp moved into a per-turn message after the cached part, and a 20-message sliding window was replaced by compaction, because "Dropping the oldest message every turn shifts the whole history by one, so the prefix changes every turn by construction." They now chart the share of input tokens read from the cache. "The ratio is the one we watch," and "It also doubles as a regression alarm."