DEV Community

Cover image for Which Claude model should you use?
Lukas Polak
Lukas Polak

Posted on Edited on

Which Claude model should you use?

Start with Opus

You should start with the Opus model at its default medium effort level. It costs half as much as Fable and is selected by default, so it's recommended for most day-to-day tasks. Opus is also a great model to start with for vision-heavy tasks.

When to move down

Move down to Sonnet for well-scoped features and small bug fixes. However, even if a task is well-scoped, that doesn't mean Sonnet is the best model for it. Major refactors or tricky bugs should be handled by larger models. Sonnet models are great for tasks you run repeatedly, such as investigations, reviews, or creating drafts.

Move down to Haiku for high-volume tasks, such as clarification or routing.

Small models are usually great for high-volume tasks, but they're less suited to long agentic loops. Use them where mistakes are easy to spot and quick to fix.

When to move up

Move up to Fable for the trickiest technical problems. Before moving up to Fable, though, you could try Opus at high effort. Keep in mind that adjusting the effort level beyond high can cost as much as using a larger model.

If you move up one model, use a lower effort level than you used with the smaller model.

A higher effort level helps when the model needs to think longer, call more tools, or check its own work, but it doesn't give the model knowledge or skills it doesn't have.

Ultracode and subagents

For fairly complex tasks, you could enable ultracode. It doesn't change the model; it just changes how Claude organizes the work. It will automatically plan a workflow for substantial tasks and coordinate subagents when useful.

Subagents run in their own context windows and hand back only a summary. The main model keeps the thread.

Price, speed and latency

When choosing a model, price isn't the only trade-off. Bigger models cost more and are more capable, but they have higher latency. Effort level and latency are roughly correlated.

Opus and Fable models have fast mode, which lowers latency without changing their quality or capabilities. One drawback is that it costs more, so be mindful of that.

A higher price per token doesn't necessarily mean you'll get better results.

Self-verification with tests or scripts using smaller models could be more cost-effective and quicker than using a bigger model.

Next-generation models are usually not only faster (some by 30%) but also cheaper and more capable.

Cache

There is no read-cache price difference between Opus and Fable models, but writing to the cache costs twice as much.

Switching models does not transfer the cache; it has to be built from scratch.

11% of this text is AI

Top comments (0)