Which Claude Model to Use: Haiku, Sonnet, Opus or Fable

People ask me which Claude model they should use, and they usually expect a ranking. The honest answer is less exciting. Most of the time the brief you write matters more than the model you pick. But picking badly still costs you something: either you wait too long for simple work, or you burn through your limits on tasks a lighter model would have handled fine.

Here is how I think about the lineup as of September 2026, and the small set of habits that keep the choice from becoming a daily debate.

The four models right now

Anthropic's models overview lists four current models. In plain terms:

• Haiku 4.5 is the fastest and cheapest. Anthropic describes it as the fastest model with near-frontier intelligence. It is great at small, clear jobs where speed matters more than depth.
• Sonnet 5 is the balanced one: fast, capable and much cheaper than the bigger models. It handles writing, analysis and everyday coding well.
• Opus 5.5 is built for long running agentic coding and knowledge work. This is the model that stays on track across a big, messy task.
• Fable 5.1 is the most capable of the four, aimed at demanding reasoning and long horizon agentic work. It is also the slowest and the most expensive.

Older versions like Opus 4.8 and Sonnet 4.6 are still around as legacy models. Unless you have a reason to stay on one, use the current generation.

The app and the docs give different defaults

This confuses a lot of people, so it is worth saying clearly. The Claude Academy guide for the Claude app tells you to start with Sonnet if you are unsure. The developer docs say that if you are unsure, start with Opus 5.5 for most workloads. In Claude Code, the default model on paid plans is Opus 5.5.

Both are right, because the cost works differently in each place. In the app you pay with your usage limits, and Sonnet stretches them much further. When you build software or run agents, a wrong answer costs you hours of cleanup, so the stronger default pays for itself.

My simple version: chatting and writing in the app, start on Sonnet. Building things in Claude Code, stay on Opus. Reach for Fable only when you can name why the job is harder than usual.

A quick way to pick

Ask these in order and stop at the first yes.

1. Will this run for a long time on its own, across many steps? A full feature build, a deep research report, a migration across a large codebase. Use Opus 5.5, and move to Fable 5.1 if Opus at a higher effort still falls short.
2. Is it a hard call that is expensive to get wrong? Architecture, pricing logic, a tricky bug you have already failed to fix twice. Use Opus 5.5.
3. Will a person read the output, and does the wording matter? Emails, product copy, a proposal. Sonnet 5 is excellent here.
4. Is it mechanical, repetitive or high volume? Tagging, reformatting, pulling fields out of text, summarizing a thread. Haiku 4.5 does it in a blink.

If none of these clearly applies, use your default and move on. Overthinking the choice costs more than a slightly wrong pick.

Effort is the dial most people forget

Current models think adaptively, and you can steer how hard they think. Anthropic's own guidance says tuning effort is often a better lever than switching models. In the Claude app, the effort setting lives in the model picker. In Claude Code you can change it for the whole session:

/effort            # open the effort slider
/effort high       # set a level directly

Or ask for deeper reasoning on just one message by including the keyword ultrathink anywhere in your prompt. The model configuration docs are explicit that only that exact word is recognized. Phrases like "think hard" are treated as normal text.

The checkout total is wrong when a coupon and a gift card are combined.
Find the root cause before changing anything. ultrathink

In practice I keep one model for the session and raise effort for the one or two hard moments, instead of hopping between models all day.

How to switch models

In the Claude app, choose the model from the model picker in the chat. In Claude Code:

/model                 # interactive picker
/model sonnet          # switch by alias
claude --model fable   # start a session on a specific model

The aliases haiku, sonnet, opus and fable always point to the latest version of each family. There is also opusplan, which uses Opus while you are in plan mode and Sonnet when it starts executing.

One trick I use a lot: give lightweight helper agents a cheaper model. A subagent that only scans files or collects facts does not need your biggest model. Add one line to its definition file in .claude/agents/:

---
name: file-scout
description: Finds where things live in the codebase and reports file paths
tools: Read, Grep, Glob
model: haiku
---
Search the codebase and report the relevant files with one line on each.

What it costs

If you use the Claude app on a subscription, you do not count tokens. You get usage limits, and heavier models use them up faster. On the Free plan you get Haiku and Sonnet. Fable is available on paid plans, but how it counts depends on the plan: Max includes it within part of your weekly limits, while Pro uses paid usage credits.

If you build on the API, these are the base prices per million tokens (input, then output), from the pricing page:

• Haiku 4.5: $1 and $5
• Sonnet 5: $2 and $10
• Opus 5.5: $4 and $20
• Fable 5.1: $10 and $50

Output always costs more than input, so long answers are where the money goes. Asking for shorter answers is often the cheapest optimization you can make.

Mistakes worth avoiding

• Using the top model for everything. It feels safe. In reality you wait longer for simple work and hit your limits sooner.
• Blaming the model for a vague brief. If the output is generic, a bigger model usually gives you a more polished version of the same generic answer. Fix the prompt first.
• Switching models mid task to "try again." Change the effort or the instructions first. Switching resets your intuition about what is going wrong.

Pick one default for a week and notice the moments where it clearly struggles or clearly wastes time. Those moments tell you when to step up or down far better than any comparison chart.