Model switches and the prompt cache
Does switching Claude models mid-session waste the prompt cache? It can. The check AutoMode runs before every model switch, and when it holds instead.
Each model keeps its own prompt cache. A switch in a warm chat writes the whole chat to the new model's cache. The effortless README reports about 5 times the cost of staying on a warm Opus in an 80k-token chat. For that reason, effortless ships model switching off.
AutoMode switches models by default, so it checks each switch first. The check is in hooks/lib/policy.ts. The route holds unless the switch passes.
| Situation | Rule |
|---|---|
| Risk floor | A mandatory floor always switches, whatever it costs. Quality comes first. |
| No warm cache | At session start, after /clear, or on a cold cache, there is little to lose, so a switch is allowed. |
| Moving up | Needs two votes in a row, unless the chat is under 20k tokens. |
| Moving down in a warm chat | No risk flagged, and judge confidence of at least 0.85. |
| Moving down in the same task | At least 8 requests since the last switch, and a passed check. |
| Moving down in a chat of 20k tokens or more | The cautious estimate of the saving must exceed 1.5 times the cache rewrite. |
| Too many switches | After 3 discretionary switches in one task, the route holds. |
Effort is free to change
On the Anthropic API, a change of effort keeps the cache for the 5.5 models and Fable 5.1. So AutoMode changes effort on every prompt and holds the model. Where a thinking setting could make the API refuse a per-message effort change, such as MAX_THINKING_TOKENS=0, AutoMode leaves the effort as Claude Code set it. AUTOMODE_THINKING_GUARD=off or on overrides that guard.
When a task fails mid-way
Under the default strategy, v6c, AutoMode does not move a large failed conversation to a pricier model. A fresh builder takes over instead, with a brief of the task, the corrections and what already failed. In an earlier version, six such switches cost $8.18 more than Sonnet high on the same tasks.
The countdown
The bar shows cache 58:12: the time until the prompt cache goes cold. After that, the next message pays full price to re-read the chat. See What the bar shows.
What it adds up to
These gates and the fresh builder are why AutoMode costs 51% less than Opus in mixed sessions. They do not make it cheaper than Sonnet high on long single coding tasks, where it costs about 15% more. See Benchmark results.
To route effort only and keep the model you picked, run /plugin configure automode@automode and set Route models to off.
Last updated on
The guardian
The AutoMode guardian redacts secrets, pauses turns that stop converging, shortens long output, saves checkpoints and guards subagents. Its five modes.
Route learning
How AutoMode learns per project: a kind of task that needed rescue starts on a stronger model next time. It only raises a route and stores no prompt text.