Grok-Bot cache bug burning quota

Where does the bug appear (feature/product)?

Grok Bot

Describe the Bug

At 2026-09-24 06:03 UTC, grok-bot-default usage rows change in two
ways at the same moment:

  1. Almost every row starts reporting an explicit cache-write column
    (previously almost all input was “without cache write”).
  2. Cache no longer survives a gap longer than about five minutes.

Same traffic before and after that timestamp:

  • Gap 5–60 minutes: before 06:03 UTC most turns still hit cache;
    after 06:03 UTC almost every turn is a cold rewrite (cache read 0,
    full context written again).
  • Gap under ~5 minutes: hit rate is almost unchanged.
  • After 06:03 UTC a repeated row shape appears: cache read exactly
    565 tokens, then a large cache write. Only a 565-token prefix is
    reused.

This is not the usual “each turn re-reads the conversation.” That
already appears as cache read. After 06:03 UTC a quiet 15–60
minute gap (routine, poll, or idle chat) becomes an uncached
rewrite of the whole context.

The timestamp is the morning after the 2026-09-23 Cursor x post that cache
reuse had improved and that those harness changes were being
applied to Grok Bot.

This increases the quota usage for the same input/output tokens by roughly 12x.

Steps to Reproduce

  1. Use a grok-bot-default session that continues across gaps of
    5–60 minutes (scheduled routine or a chat left idle).
  2. Export All Events for grok-bot-default.
  3. Split rows at 2026-09-24 06:03 UTC.
  4. For each side, compute cold-start rate by gap since the previous
    grok-bot-default row (cache read = 0, or uncached ≈ full prompt).

Expected Behavior

A 5–60 minute gap still reads the existing prefix, as it did before
06:03 UTC.

Operating System

MacOS

Version Information

Version: 0.61.0
Built: 2026-09-26T21:54:55.069Z
Release Track: stable
OS: darwin

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor