When to switch Claude models
Switching models in the middle of a chat throws away the cache, so the next message rereads everything at full price. Here is when to switch, and how the pro avoids it.
By Frontpage AI · Checked against Anthropic's docs on · From our Reel "Part two. Same three people. Only one switches Claude like a pro." Part one: which model to use.
The short answer
Pick your model at the start of a chat and keep it. Each model has its own cache, so a mid-chat /model switch makes the next message reread the whole chat at full price, 20 to 40 times what a cached read costs. New job? Type /clear. Keep simple helpers on Haiku 5.5 with model: haiku. Need two models at once? Run each in its own terminal with claude --model.
What are the three ways people switch Claude models?
Beginner
Hops models in the middle of a chat
Opus for one message, Fable for the next, back to Sonnet for a quick fix.
Every hop rereads the whole chat at full price.
Improver
Picks the model at the start, clears the chat for each new job
The cache stays warm. But switch to Fable for the hard job, and the helpers Claude hands side jobs to switch with it.
Simple jobs like reading logs end up running on Fable.
Pro
Never switches. One terminal per model.
Quick jobs go to a terminal on Sonnet 5.5. The hard one gets its own terminal on Fable 5.2. Simple helpers run on Haiku 5.5.
Each cache stays warm, and the simple work runs up to 80 times cheaper than Fable.
Why does switching models mid-chat cost more?
Every time you send a message, Claude rereads the whole chat so far. Claude Code keeps that chat in a cache, so the reread is cheap. But each model has its own cache. In Anthropic's words:
"Switching with /model means the next request reads the entire conversation history with no cache hits, even though the content is identical."That's why Claude Code asks you to confirm a /model switch while the cache is still warm. One switch costs you once, for the whole chat so far. Hopping every few messages costs you every time.
| Model | Full read | Cache read | Cache saves |
|---|---|---|---|
| Fable 5.2 | $8 | $0.20 | 39 in 40 |
| Opus 5.5 | $4 | $0.20 | 19 in 20 |
| Sonnet 5.5 | $2 | $0.10 | 19 in 20 |
| Haiku 5.5 prompts up to 100,000 tokens | $0.10 | $0.01 | 9 in 10 |
These don't reset the cache
Turning the advisor on or off. On current models, changing /effort. And opusplan from part one, which switches once per plan, when you leave plan mode.
How long the cache lasts
On a Claude subscription, within your plan's usage, an hour since your last message. With an API key, five minutes, unless you set promptCacheTtl to 1h.
When should you switch, then?
Anthropic's advice: "Pick your model and effort level at the top of a session, then save /compact for natural breaks between tasks. The fewer changes you make mid-task, the higher your cache hit rate."
Starting a chat
Pick the model nowChoose with /model opus, or start with claude --model opus, and keep it for the whole task.
A new job that has nothing to do with the last one
Type /clearAnthropic: "When you want a fresh start instead of continuity, /clear costs nothing." Pick the model for the new job straight after.
You really do need a different model for this chat
Switch at a breakDo it once, between tasks, not every few messages. The first message after the switch is slower and dearer. After that, the new cache builds up again.
A hard job and quick jobs at the same time
Two terminalsGive each model its own terminal, below. Nothing gets switched, so nothing gets reread.
How do you stop helpers following you to Fable?
Claude Code hands side jobs, like research and running tests, to helpers called subagents. Unless you set their model, they use the one your chat is on. Anthropic: "Switch to Opus before Claude delegates research or test runs to one of them, and that work runs on Opus too." The same goes for Fable.
To keep a simple helper small, give it its own file with model: haiku. Put it in .claude/agents/ for this project, or ~/.claude/agents/ for all your projects:
---
name: log-reader
description: Reads log files and reports the errors, with the file and line.
model: haiku
---
Read the logs you're pointed at. List each error with its file, line and message. Don't fix anything.Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, against $8 and $40 for Fable 5.2. That's up to 80 times cheaper. Over 100,000 prompt tokens it's $0.50 and $2.50, still 16 times cheaper.
Which model for which helper
| Helper's job | model: | What Anthropic's docs say |
|---|---|---|
| Simple jobs: reading logs, finding files, running tests | haiku | "For simple subagent tasks, specify model: haiku" |
| Writing code from a plan | sonnet | "Sonnet handles most coding tasks well and costs less than Opus." |
| Big design calls, tricky reasoning | opus | "Reserve Opus for complex architectural decisions or multi-step reasoning." |
| The hardest, longest jobs | fable | "for your hardest and longest-running tasks" |
- Keep harder helpers, like real research or tricky reviews, on a bigger model, or on
inheritto follow your chat. - The built-in Explore helper, the one that searches your code, has its own rule: when your chat is on Fable with a subscription, it runs on Opus. To make it cheaper, add your own helper named
Explorewithmodel: haiku. It replaces the built-in one. - To put every helper on one model, set the
CLAUDE_CODE_SUBAGENT_MODELenvironment variable (for example tohaiku) in your settings. A model set in a helper's own file still wins, unless you also setCLAUDE_CODE_SUBAGENT_MODEL_FORCE=1(Claude Code v2.1.257 or later). - Type
/tasksto see which model and effort level each running helper is using.
How do you run two Claude models at once?
A quick job and a hard job don't have to share one chat. Anthropic: "To run different models in different terminals at the same time, launch each one with its own --model flag rather than switching with /model."
# Terminal 1: quick jobs
claude --model sonnet
# Terminal 2: the hard one
claude --model fableEach terminal is its own session, with its own model and its own cache, so neither throws away the other's. The --model flag only applies to the session you start with it, and doesn't change your saved default.
Quick reference
| Do this | Command |
|---|---|
| Choose the model before you start | /model opus or claude --model opus |
| Start a fresh chat for a new job | /clear |
| Keep simple helpers small | model: haiku in .claude/agents/ |
| Run two models at once | claude --model sonnet and claude --model fable |
| See what you've spent | /usage |
How much does each Claude model cost?
| Model | Input | Output | Cache read |
|---|---|---|---|
| Fable 5.2 | $8 | $40 | $0.20 |
| Opus 5.5 | $4 | $20 | $0.20 |
| Sonnet 5.5 | $2 | $10 | $0.10 |
| Haiku 5.5 prompts up to 100,000 tokens | $0.10 | $0.50 | $0.01 |
| Haiku 5.5 prompts over 100,000 tokens | $0.50 | $2.50 | $0.05 |
Questions people ask
Does switching models in Claude Code reset the prompt cache?
Yes. Each model has its own cache. After you switch with /model, the next message reads the entire conversation again with no cache hits, at the new model's full input price. After that, the new model's cache builds up again. This is why Claude Code asks you to confirm a /model switch while the cache is still warm.
How much cheaper is a cached read than a full read?
On the Claude API, a cache read costs a twentieth of the normal input price on Opus 5.5 and Sonnet 5.5, for example $0.20 against $4 per million tokens on Opus 5.5. On Fable 5.2 it is a fortieth: $0.20 against $8.
Does changing effort or turning the advisor on or off reset the cache?
No. Turning the advisor on or off keeps the cache, and on current models changing /effort does too. Switching the main model with /model is what resets it. opusplan switches between Opus and Sonnet once per plan, when you leave plan mode, which is a natural break and fine.
Do Claude Code subagents use the same model as my chat?
Yes, unless you set their model. Switch your chat to Fable and helpers doing research or running tests run on Fable too. Add model: haiku to a simple helper's file in .claude/agents/ to keep it on Haiku 5.5, which is up to 80 times cheaper than Fable 5.2.
How do I run two Claude models at the same time?
Open two terminals and start each one with its own model: claude --model sonnet in one and claude --model fable in the other. Each terminal is its own session with its own cache, so neither throws away the other's. The --model flag only applies to that session and does not change your saved default.
How long does the Claude Code prompt cache last?
On a Claude subscription, within your plan's usage, it lasts an hour since your last message. With an API key it lasts five minutes, unless you set promptCacheTtl to 1h.
Sources
- Claude Code: prompt caching (switching models, cache lifetime)
- Claude Code: manage costs (/clear, model: haiku for helpers)
- Claude Code: model configuration (helpers follow /model, --model per terminal, model aliases)
- Claude Code: subagents (Explore's model, CLAUDE_CODE_SUBAGENT_MODEL, file format)
- Claude API pricing
Model names, prices and commands change. Check the sources before relying on them. Last checked .
Run a business? See what AI says about you.
Frontpage AI builds websites for independent UK service businesses that ChatGPT, Gemini and Google can read and recommend. We run a free 15-minute AI check, whether or not you become a partner.
15 minutes. No obligation. An honest answer either way.