Which Claude model is best? Opus vs Sonnet vs Haiku vs Fable

Anthropic now ships four current Claude models, and the right one depends less on which is "smartest" than on how hard the task is, how fast you need the answer and how many times you run it. This guide compares them on what matters in practice and ends with a simple way to choose.

Model names, prices and limits below are from Anthropic's model documentation, checked on 12 October 2026. For the full price list, including caching and batch discounts, see Claude API pricing.

The short answer

The current Claude models compared

* Haiku 5.5's price applies to prompts up to 100,000 tokens; above that it is $0.50 / $2.50.

All four use tools and support adaptive thinking, where the model decides how long to think. Depth is set with an effort level from low to max. Opus 5.5 and Haiku 5.5 default to medium; Fable 5.1 and Sonnet 5.5 default to high. Effort is often a bigger lever than model choice: a strong model at lower effort can beat a weaker model at high effort and cost less.

Claude Fable 5.1: the most capable

Fable 5.1 sits above Opus. It is built for the most demanding reasoning and long-horizon agent work: a large refactor across a codebase, a hard bug nobody has found, a research task that takes many steps.

Three things set it apart in practice:

• It always thinks. Thinking cannot be switched off; you control depth with effort. A single hard request can run for many minutes.
• It costs the most. $10 in and $50 out per million tokens, two and a half times Opus 5.5.
• It needs 30-day data retention. It is not available under zero data retention unless Anthropic expressly authorises it. For companies with strict retention rules, that alone can rule it out.

Use it where a better answer is worth real money: architecture decisions, hard debugging, work you would otherwise give to your most senior engineer.

Claude Opus 5.5: the default

Opus 5.5 is the model Anthropic recommends by default, and the right starting point for most teams. It is strong at multi-step reasoning, coding and agentic work, and at $4/$20 it is cheaper per token than the Opus 5 it replaced ($5/$25).

It also has a fast mode on Anthropic's own API: the same model at up to 2.5 times the output speed, for $8/$40 per million tokens.

Claude Sonnet 5.5: the everyday model

Sonnet 5.5 gives most of the capability at half Opus's price ($2/$10) and answers faster. It is the sensible choice for everyday coding, writing, summaries and analysis where you notice speed and the task is not unusually hard, and for products that call the model many times per user.

Claude Haiku 5.5: high volume at a low price

Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens: forty times cheaper than Opus 5.5 on input. It still has a 1M-token context window. Use it for classification, extraction, routing, tagging, summarising large volumes and as the worker model under a more capable orchestrator.

Which Claude model is best for coding?

For coding, the answer depends on how the model is used:

• Interactive coding in Claude Code or an IDE: Opus 5.5. It plans better and makes fewer wrong turns on multi-file changes, which saves more time than Sonnet's speed.
• Long autonomous runs and the hardest bugs: Fable 5.1, if the budget allows. It is the model to escalate to when Opus gets stuck.
• Quick edits, tests and boilerplate: Sonnet 5.5. Fast, cheap and good enough.
• Subagents doing narrow jobs (searching a codebase, summarising files, running checks): Haiku 5.5 or Sonnet 5.5.

A common setup is Opus 5.5 for the main session and planning, with Sonnet or Haiku for subagents. Our Claude Code guide shows how to switch models, and Claude Code pricing covers what a day of agent work costs.

Opus vs Sonnet: when is Opus worth double?

Opus 5.5 costs twice what Sonnet 5.5 does per token. It earns that when:

• the task has many steps and an early mistake is expensive (agents, migrations, multi-file refactors);
• the problem needs real reasoning, not recall (debugging, design trade-offs, analysis with conflicting data);
• a wrong answer costs more than the tokens (legal, financial or customer-facing output that someone will act on).

For one-shot questions, drafting, summaries and high-volume pipelines, Sonnet 5.5 or Haiku 5.5 usually gives the same result for less. Measure on your own tasks: run twenty real examples through both and compare cost per correct result, not cost per request.

Which models you get on each Claude plan

In the Claude apps, the plan decides the models. Free gets Sonnet and Haiku. Pro and Team Standard get Opus, Sonnet and Haiku, with Fable only through paid usage credits. Max and Team Premium include Fable up to half of the weekly limit, and Enterprise bills every model at API rates. The full comparison is in Claude plans compared.

Older Claude models

Anthropic still serves earlier models, including Claude Fable 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, Sonnet 4.6 and Haiku 4.5. They are worth keeping only if you have a tested setup that depends on them. New work should start on the current generation, which is as cheap or cheaper: Opus 5.5 costs less than Opus 5, and Haiku 5.5 costs a tenth of Haiku 4.5 ($1/$5) with five times the context.

Choosing for a company, not a person

A team rarely needs one model. It needs the right model per task, and a way to stop everyone defaulting to the most expensive one:

• Route by task. Opus for planning and hard work, Sonnet for the everyday, Haiku for bulk jobs.
• Set budgets per team or user, so a long Fable session is a decision, not an accident.
• Check data handling per model. Fable 5.1's retention requirement, for example, matters if you have zero-retention commitments.
• Keep the choice reversible. Models change every few months; a gateway lets you switch without touching every integration.

Walma AI Hub gives every employee Claude, GPT and other models through one gateway in your own EU region, with per-team model access, budgets and a full log. See how it works, or read about running Claude in the EU.