Every follow-up paid for the whole conversation again.
At 100,000 follow-up turns a month, that is about $1,700 to $5,100 a month, depending on how long conversations run.
Metabot caches well inside a turn. Across turns it didn't, because the conversation it sent back was not exactly the one it had sent before. A context block with a date to the microsecond wasn't saved, thinking blocks were dropped, parallel tool calls were split, and a failed tool result lost its error flag.
So the first call of every follow-up turn read only the tools and system prompt (12,600 tokens) from the cache and wrote the rest of the conversation again, at the 1.25× cache-write price.
The fix is to replay each turn exactly as it was sent, so every turn starts with everything the last one cached. That changes how conversations are stored, which is their team's call, so we reported it as an issue with the evidence and offered our patch.
| turn | saved | per 1,000 turns |
|---|---|---|
| 2 | 21–31% | ≈ $17 |
| 3 | 46% | ≈ $28–44 |
| 4 | 56–64% | ≈ $33–51 |