Switching models simply means choosing a different Claude model to answer your next message.

When you open Claude and start typing, a particular model is already selected. That model reads your message and writes the reply. If you decide that a different model would suit your task better, you can change the selection from a menu near the send button. From that point on, the new model answers your messages. You do not have to start over, create a new account, or open a different website.

This is possible because Anthropic, the company that makes Claude, offers more than one model at the same time. They are all “Claude”, they share the same general character and values, but they differ in how fast they respond, how deeply they can reason, how much of your usage allowance each answer consumes, and which tasks they handle best. Switching is how you take advantage of those differences.

Switching can happen in two ways

The first way is manual, which means you decide and click. This is the main topic of this guide. The second way is automatic, which means Claude itself moves a request to a different model in certain special situations, and tells you when it does. We cover both, because understanding the automatic kind prevents confusion when you see a notice you did not expect.

A handful of simple words appear throughout this topic. Here they are in plain English.

Four families make up the main lineup, each aimed at a different point on the speed, depth, and cost scale.

At the time of writing, the models most people can choose from are Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, and Claude Fable 5.1. The names come from forms of writing. A haiku is short and quick, a sonnet is balanced and structured, and an opus is a large, serious work. Fable sits above Opus as a higher tier built for the most demanding work. There is also a separate Mythos tier, but it is not broadly open to the public and is used by a small number of trusted organizations.

There are a few clear situations where changing models pays off.

In the Claude apps the process takes only a few seconds.

flowchart LR
    A["Open or continue a chat"] --> B["Click the model name by the send button"]
    B --> C{"Is your model in the short list?"}
    C -- "Yes" --> D["Select it"]
    C -- "No" --> E["Choose More models"]
    E --> D
    D --> F["Send your next message"]
    F --> G["New model writes the reply"]
        

The same idea on web, desktop, and mobile

Claude is available on the web, as a desktop app, and as a mobile app. The details of the screen layout can differ a little between them, but the idea is the same everywhere: look for the model name beside the message box and tap or click it. If you cannot find it, the official Help Center article called “Change the model, effort, and thinking settings” explains the current layout.

A common worry is losing your work. In practice, the chat simply continues.

Switching models does not wipe the conversation. The messages already on the screen stay where they are, and your next message is simply answered by the model you just selected. Because the new model is continuing the same chat, it can build on what has already been discussed instead of starting from a blank page.

The change is also forward-looking. Replies that were already written by the earlier model are not rewritten. Only new replies come from the new model. This means a single conversation can contain answers from more than one model, which is perfectly normal.

Things worth keeping in mind

Choosing the model is only half of the story. You can also decide how hard the model should work.

Besides the model itself, the same menu lets you pick an effort level for models that support it. Effort controls how much thinking Claude applies to each response. Higher effort gives more thorough answers but takes longer and uses more of your allowance. Lower effort is quicker and cheaper. So you really have two dials: which model, and how much effort.

The effort selector is offered for the newer models in the lineup, including Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1, along with a number of earlier versions. Each model has a recommended starting level, marked “Default” in the menu, and for most everyday tasks that default works well.

flowchart TD
    A["Your task"] --> B{"How hard is it?"}
    B -- "Simple" --> C["Lighter model, low or medium effort"]
    B -- "Everyday" --> D["Balanced model, default effort"]
    B -- "Complex" --> E["Stronger model, high effort"]
    B -- "Extremely hard" --> F["Top model, extra high or max effort"]
        

Thinking and effort sound alike, but they are two separate settings that work together.

Thinking lets Claude spend extra time breaking a problem into pieces, planning, and exploring approaches before it writes the final reply. When thinking is on, you may see a “Thinking” indicator with a timer, and an expandable section above the answer where you can read a summary of how Claude approached the problem. Reading that section can help you check how it reached its conclusion.

Effort controls how thorough Claude is in every response. The thinking toggle controls whether Claude works through its reasoning in a visible section before answering. You can combine them in any way. For example, you might choose a medium effort with thinking on, or a high effort with thinking off, where that choice is allowed.

Not every model lets you turn thinking off

On some of the newest models, thinking is built in and cannot be switched off in the Claude app. This applies to Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1, among others. On Opus 5.5 and Fable 5.1, thinking is always on at every effort level. If you do not see a toggle on a model, this is usually the reason, not a fault.

Where to find it

Here is a practical guide that links common tasks to a sensible starting choice.

Treat this table as a starting point and not a strict rule. If an answer disappoints you, the easiest experiment is to change one thing at a time. Try the same question on a stronger model, or raise the effort level, and compare the results.

flowchart TD
    A["Start a task"] --> B["Begin on a balanced model"]
    B --> C{"Is the answer good enough?"}
    C -- "Yes" --> D["Continue as you are"]
    C -- "Too shallow" --> E["Raise effort or move to a stronger model"]
    C -- "Too slow or too heavy" --> F["Lower effort or move to a lighter model"]
    E --> C
    F --> C
        

Sometimes you will see a notice that your chat moved to a different model. This is automatic switching, also called fallback.

Anthropic builds safety checks into its most capable models. These checks look at the request and at everything the model reads, including memory, content from connected apps, web search results, and files. For a narrow set of higher-risk requests, the chat does not simply stop. Instead, the request can be re-run on a different, earlier model so that everyday work can continue while risks are limited.

Take Claude Sonnet 5.5 as an example. For certain higher-risk cybersecurity requests, such as exploit generation or penetration testing, and for a small set of capabilities tied to developing the most advanced AI systems, a request can fall back to Claude Sonnet 5. Most requests never trigger any of this. Routine software development and most life-science work are unaffected.

sequenceDiagram
    participant You
    participant Claude as Selected model
    participant Check as Safety check
    participant Backup as Fallback model
    You->>Claude: Send a message
    Claude->>Check: Request is screened
    Check-->>Claude: Flagged as higher risk
    Claude->>Backup: Re-run in the same chat
    Backup-->>You: Reply with a notice that the model switched
        

Some requests are blocked instead of switched

Not every flagged request falls back. Certain categories, such as requests that could help cause serious biological harm, and attempts to extract the model’s internal reasoning word for word, are blocked outright instead of being moved to another model. You can still ask Claude to explain its reasoning in a normal conversational way, teach you a concept, or walk through a review. Simple questions such as “why did you do that?” are not affected.

You can control automatic switching

Automatic switching is on by default for models that use it. You can turn it off from the settings area under Capabilities by switching off the option that moves models when a message is flagged. When it is off, a request that would have fallen back pauses the chat instead. You can then edit your message and try again on the original model, or send the same message to a different model yourself.

If a model or effort level you expected is missing, there is usually a simple explanation.

Claude appears in several places, and each one has its own way to choose a model.

The key lesson is that the idea of matching a model to a task is the same everywhere, even though the buttons and commands differ. When you move to a new surface, look for the official page for that product to learn the current steps.

Every model choice is a trade between three things you care about.

You can think of model selection as a triangle with quality, speed, and usage at its corners. A stronger model gives higher quality but is slower and uses more of your allowance. A lighter model is faster and lighter on usage but may give less depth on very hard problems. Effort level slides you along the same triangle, because raising it improves thoroughness and lowers speed and allowance together.

A few habits waste time, usage, or quality. Here is how to avoid them.

These simple habits make model switching feel natural and effective.

Let us follow an imagined person as they prepare a presentation for their team.

In one conversation she used three different settings, each chosen to fit the step she was on. She saved time on easy parts, got depth on the hard part, and protected herself with a final check. That is the whole spirit of switching models.

The list of models is not frozen. New versions arrive and older ones are retired.

AI models improve quickly. A model that is the newest today may be replaced by a better version in a few months. The family names such as Haiku, Sonnet, Opus, and Fable tend to stay as familiar labels, while the numbers attached to them rise. A new version of a family generally keeps the same role, for example Sonnet remains the balanced everyday choice, but becomes more capable.

Because of this, the most useful skill is not memorizing which version is best. It is understanding the roles. If you know that one family is fast and light, another is balanced, another is deep, and another is the very top tier, you can make a good choice from whatever menu you see, even after the version numbers change.

flowchart LR
    A["New models are released"] --> B["Family roles stay familiar"]
    B --> C["You learn the roles, not just the numbers"]
    C --> D["You choose well from any menu"]
        

Quick answers to the questions beginners ask most often.

Yes. You can change the model, the effort level, or the thinking setting at any point in a chat. The change applies starting with Claude’s next response.

No. The conversation stays in place and simply continues with the new model. Earlier replies remain as they were.

It sits next to the send button in the message area. It shows the current model and effort level, and clicking the model name opens the list.

The model decides which Claude is answering. The effort level decides how much thinking that model puts into each reply. They are two separate dials you can combine.

On several of the newest models, thinking is built in and cannot be disabled in the Claude app. On some of them it is always on at every effort level.

A safety check may have flagged a higher-risk request, so Claude re-ran it on a fallback model and showed you a notice. You can switch back manually, or turn automatic switching off in settings under Capabilities.

Check “More models” first. If you use an organization plan, your administrator may have turned that model off for your role. Updating the app is also worth trying.

Not always. For simple tasks, a lighter model can be just as good and much faster. The strongest model earns its extra cost on hard, complex, or high-stakes problems.

It can. Heavier models and higher effort levels use more of your allowance per reply, so you reach your limit sooner. Lighter models and lower effort stretch it further.

The idea is the same, but the steps differ. Claude Code and the API have their own settings and documentation, and automatic fallback must be set up by developers on the API.

Start with the balanced everyday model at its default effort. Move up if the answer is too shallow, and move down if you want more speed or want to save usage.

Yes. Every model can make mistakes. Keep verifying important facts, and check more carefully when the consequences of an error are serious.

Claude is a family of models, not a single one, and you can change which model answers you at any moment from the menu beside the send button. Haiku is quick and light, Sonnet is the balanced everyday choice, Opus is built for deeper work, and Fable sits at the top for the hardest problems. Alongside the model, the effort level and the thinking setting let you control how thorough each reply is. Changes apply from the next response, the conversation continues without loss, and in a small number of higher-risk situations Claude may move a request to another model automatically and tell you so. The best approach is to start balanced, adjust one setting at a time, and always match your choice to the difficulty and the stakes of the task.