Choosing a Claude Model: Fable, Opus, Sonnet or Haiku

1.Four factors and four tiers

Every model choice trades quality against latency and cost. Anthropic asks you to weigh four factors: the capabilities the task needs, how fast the application must respond, your budget for development and for production, and effort. Effort is a parameter on several models that trades intelligence for latency and cost without changing the model.

The current models spread across that range. All of them handle text and image input, text output, multiple languages, vision and tool use. They differ in speed, price, thinking mode, context window and maximum output.

By default, start with Claude Opus 5.5. Move to Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 still falls short on your evals at higher effort. The documented use cases for each tier:

- Fable 5.1: agent sessions that run for hours, and multistep deep research. - Opus 5.5: multihour autonomous coding agents, large-scale refactoring and computer use. - Sonnet 5: everyday code generation, data analysis and content creation. - Haiku 4.5: the lowest latency and price, for real-time applications, high-volume processing and sub-agent tasks.

For latency there is also fast mode. Claude Opus 5.5, Claude Opus 5 and Claude Opus 4.8 support it as a research preview, which gives up to 2.5x higher output speed at premium pricing.

Sources12

2.Cost-first or capability-first

The docs describe two ways to start. Which one fits depends on how hard the task is and on what a wrong answer costs.

The two approaches move in opposite directions. Cost-first starts cheap and upgrades only when a specific capability gap shows up. Capability-first starts strong and saves money later, first by lowering effort and then by moving to a smaller model. When accuracy matters more than cost, for example in scientific or mathematical work, the guidance points to capability-first.

Sources2

3.Adaptive vs. extended thinking by model

Thinking lets Claude work through a problem before it answers. It restates the question, tries approaches, checks intermediate results and drops paths that don't work. That helps on tasks where the quality of the answer depends on intermediate work. Thinking also costs money: reasoning tokens are billed as output tokens and count toward max_tokens.

The current models don't all handle thinking the same way:

- Adaptive thinking lets Claude decide when to think and how deeply. - Extended thinking is configured by hand with a fixed token budget. Haiku 4.5 is the current model that uses it. - On some models thinking is always on. On others it is on by default and can be turned off, or off until you ask for it.

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=4096,
    thinking={"type": "disabled"},
    messages=[{"role": "user", "content": "Summarize this article in one sentence."}],
)

The display setting also affects latency. On Opus 5.5, Sonnet 5 and Fable 5.1, display defaults to "omitted", so thinking blocks come back with an empty text field. When you stream, this gets the first text token out sooner, because the server skips streaming the thinking tokens. You pay for them either way.

Sources3

4.Letting evals make the decision

Tables and use-case lists only give you a starting point. To decide whether to upgrade or change models:

1. Build a benchmark for your own use case. 2. Run it with your real prompts and data. 3. Compare the models on accuracy, response quality and edge cases. 4. Weigh that performance against cost.

Without a good eval set, you can't tell whether a cheaper tier holds up or a pricier one is worth it.

Sources2

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Every current Claude model uses adaptive thinking and accepts an effort level.Why is that wrong?

   Claude Haiku 4.5 uses extended thinking, set with type "enabled" and budget_tokens, and does not support effort. Fable 5.1, Opus 5.5 and Sonnet 5 use adaptive thinking.

   Covered in Adaptive vs. extended thinking by model

2. 2.If thinking text is hidden (display "omitted"), the thinking is free.Why is that wrong?

   Thinking tokens are billed as output tokens whether or not the text is returned. Omitting it speeds up the first streamed text token but does not reduce cost.

   Covered in Adaptive vs. extended thinking by model