Chats and what they cost¶
Warm and cold¶
Claude keeps a chat cheap while its cache is warm: the next message reads the chat from the cache at a tenth of the price. Claude Code's cache lasts 5 minutes or 1 hour after each reply. Once it runs out, the chat has gone cold, and the next message reads the whole chat again at full price.
Switchboard measures each chat's cache from Claude Code itself (its status line, or how long the chat's cache has lasted before), and never guesses. A card says Cache good for 4:12 while it's warm; a chat whose cache isn't known yet says cache ?.
| Size | Cold, going back... |
|---|---|
| Under 60k tokens | is cheap anyway. Resume |
| 60k to 150k | reads it all again. Start fresh is cheaper if you don't need the whole history |
| 150k or more | reads it all again at full price, and the guard holds your message first |
Resume or start fresh¶
For a quiet chat, the details compare two ways back, and mark one Suggested.
- Resume goes back to the chat as it is. Warm, that's cheap. Cold, it reads every token again.
- Start fresh writes a summary of the chat on its computer, then opens a new chat in the same folder with just that summary. The summary keeps every instruction you gave, word for word, and points to the old chat in case Claude needs more. The old chat moves to Replaced, and the new one is named after it with "(part 2)".
The suggestion follows a few rules, in order: a working chat gets none (open its terminal instead); a held message suggests Start fresh; a warm chat, Resume; a chat of 270k or more, close to where Claude Code compacts it anyway, Start fresh; otherwise whichever costs less.
How are the costs estimated?
In tokens, weighted by API prices: reading from the cache counts 0.1, writing the cache 1.25 (5-minute cache) or 2 (1-hour cache), plain input 1, and output 5. Your plan may weigh them differently, so they are estimates.
- Resume: the whole chat, read from the cache (warm) or written to it again (cold).
- Start fresh: writing the summary (a condensed copy of the chat: your messages and Claude's replies, without tool output, plus about 2,000 tokens of summary) and the new chat reading that summary. When a summary written after the last reply already exists, it is used again for free.
The details' How is this estimated? link shows the numbers for that chat.
Summaries are written by Opus, or Sonnet if you pick it in Settings, New sessions ("Sonnet is quicker and costs less of your plan"). Writing one takes about a minute. If the chat is still working, Switchboard asks first: start fresh when it finishes, or now, which stops it.
The guard for big cold chats¶
A message typed into a chat of 150k tokens or more whose cache has gone cold is held, so it doesn't read the whole chat again by surprise. In the terminal you see what it would cost, and two ways on:
- Start fresh (cheaper): type
/clear, then send anything. Switchboard writes a summary, loads it in a new chat and carries your message over. - Keep going: send the same message again, or type
go, within 15 minutes.
On the board the chat moves to Needs you with Message on hold, and its details offer Send anyway and
Start fresh. Commands starting with / are never held.
Start a new chat¶
N on the board, New terminal on a computer's row, or the command palette opens New chat.
| Machine | Any computer that's online, or Auto |
| Folder | The computer's Claude folder, a recent one, any folder inside it, a new one, or (when the computer allows it) anywhere in its home folder |
| Mode | Normal, Max effort (deepest reasoning on every turn) or UltraCode (multi-agent workflows for big builds, on Opus at extra high effort; uses your plan faster) |
| Model | Default, or the models your computers' Claude Code takes |
| Effort | Default, Low, Medium, High, Extra high |
| Permissions | Asks first, Accept edits, Plan, Auto, Bypass |
| Browser | Claude in Chrome |
| Tools | Auto, or pick tools from your library |
| Folders | Up to 8 extra folders Claude may work in |
| Other flags | Any other Claude Code flags |
| First message | What it should work on, or empty to start idle. Your draft is kept for a week |
Save setup keeps a combination of these by name; Last used brings back the last one. Leave Open its terminal checked to watch it start.
Auto¶
Auto picks the computer for you. By default it goes by rules: a computer with what the task needs (Xcode for iOS, an Android SDK, a GPU, Windows, going by words in your message), then one with a connected phone, then one that has the repository your message names, then the one with the most free memory. A chat that uses Claude in Chrome goes to its Chrome computer. Only computers that share what they can do (Settings, Machines) take part.
With Let Claude pick for Auto on (Settings, AI and your Claude setup), Claude makes the pick from your task and each computer's tools, free memory and repository names, and falls back to the rules if it can't.
A folder Claude hasn't worked in¶
Before Claude works in a folder outside the Claude folder for the first time, Switchboard asks you, the same question Claude Code would: Let Claude work in this folder? If you say yes, Claude can read and change the files in it and run commands there, in that folder and the ones inside it.
Continue on another computer¶
Continue on another machine (C, or Continue on
Names¶
A new chat is named from its first few words, or, with Name new chats on, by Claude from its first message.
Rename chat (or double-click the name, or the tab's title) changes it everywhere at once: the tab, the board,
the details, History, the iPhone, and Claude Code's own /resume list. Clear the name to go back to the automatic
one.