Claude Haiku vs Sonnet vs Opus vs Fable: Which Model to Use
A practical comparison of Claude's model lineup, Haiku, Sonnet, Opus, and Fable, covering speed, cost, and when to use each one.
What’s the difference between Claude’s models?
Claude isn’t one AI, it’s a lineup of models built for different jobs. Haiku is the smallest, cheapest, and fastest option, meant for quick, simple tasks. Sonnet is the everyday workhorse, good enough for most work when instructions are clear. Opus is the model you hand messy, open-ended, or high-stakes work to. Fable sits above all three as Anthropic’s top-tier model, built for long, demanding reasoning tasks where Claude works with more independence, but it costs significantly more to run and is usually unnecessary for typical work.
TL;DR
- Haiku is the cheapest and fastest Claude model, best suited for simple, high-volume, or low-stakes tasks where speed matters more than depth.
- Sonnet handles the bulk of everyday work well, especially tasks with clear instructions, and in testing it finished some jobs in about half the time Opus took.
- Opus is worth reaching for when a task is vague or open-ended, like building a multi-step plan from scratch, since it tends to handle ambiguity better than Sonnet.
- Fable is Anthropic’s most powerful model, intended for long, autonomous reasoning tasks, but it costs about two and a half times more than Opus on a pay-per-use basis.
- Anthropic’s own release notes state that Opus performs near Fable level on most work, which means Fable is overkill for the majority of use cases.
- Free Claude plans include Sonnet and Haiku but not Opus, while Fable access is tied to the Max and Ultra subscription tiers or paid API credits.
- A command called /advisor lets Claude consult a stronger model like Fable mid-task, giving you some of that extra reasoning power without running the whole job on the most expensive model.
How does each model actually differ in practice?
The differences aren’t just marketing labels, they show up in real task performance. In a head-to-head test across seven real workflows, Sonnet won on tasks with well-defined instructions, like producing a resource guide, finishing in roughly half the time Opus took and usually at lower cost. Opus pulled ahead on vaguer goals, such as building a 90-day plan from scratch, where the model has to make more judgment calls about structure and priorities rather than just execute clear steps.
That pattern is a useful mental model: Sonnet for execution, Opus for ambiguity. Haiku fits below both, intended for tasks where speed and cost matter more than depth of reasoning, like quick lookups or simple formatting jobs. Fable is positioned above Opus for the hardest, longest-running reasoning tasks, the kind where Claude needs to work autonomously over an extended session.
Is Opus worth using over Sonnet?
For a lot of work, no. If a task comes with clear instructions and a defined output, Sonnet tends to be faster and cheaper while producing comparable results. Opus earns its keep on tasks where the goal itself is fuzzy, where the model has to figure out structure, sequence, or priorities rather than just follow a recipe. Think of it less as “Opus is smarter” and more as “Opus is better at handling open-ended judgment calls.”
Cost and access also matter here. Opus isn’t available on the free plan, only Sonnet and Haiku are. So anyone on a free account doesn’t have the choice to begin with. On paid plans, the decision comes down to whether the task actually needs that extra judgment or whether it’s just burning budget for no real gain.
What is Fable, and when does it make sense to use it?
Fable 5.1 is Anthropic’s most capable model, designed for demanding reasoning and long tasks where Claude operates with more autonomy. It’s notable partly because of its power: an earlier version, Fable 5, was reportedly withdrawn before release because it was considered too powerful at the time. The current version, Fable 5.1, is built for the same category of heavy, extended work.
The catch is cost. Fable runs at roughly two and a half times the price of Opus on a pay-per-use basis. And Anthropic’s own release notes state that Opus performs near Fable level on most work, meaning the price premium often isn’t justified unless the task genuinely requires Fable’s extra capability.
Access to Fable also depends on your plan. On Max and Ultra subscriptions, Fable usage can consume up to half of your weekly limit, so it’s not something to reach for casually. On Pro, Fable is billed through usage credits rather than bundled into the subscription. For most people, the practical move is to let Sonnet or Opus handle the bulk of the work and use the /advisor command to bring Fable in at key decision points, getting some of its reasoning power without routing an entire task through the most expensive model.
How do tokens and context windows affect model choice?
Every current Claude model works within a 1 million token context window, which Anthropic estimates at around 750,000 words, enough to hold a large stack of documents or code at once. But the more important detail for model choice is how tokens get used. In one measured set of tasks, Claude read about 160 million tokens while writing only about 800,000, meaning for every token it generates, it may read roughly 200. That ratio matters because the files, instructions, and context you load in affect cost and speed far more than how long your prompt is.
This has a direct bearing on which model to pick. A larger, more expensive model reading through huge amounts of reference material will cost more regardless of how short the final answer is. For tasks that involve scanning large volumes of text but producing modest output, Sonnet or even Haiku may be the more efficient choice, reserving Opus or Fable for tasks where the reasoning itself, not just the reading, is the hard part.
How does effort level interact with model choice?
Separate from which model you pick, Claude also has adjustable effort levels: low, medium, high, extra high, and max. The default is medium. In one test, the same prompt was run through Opus at every effort level. Low finished in about 17 minutes for under $4 but produced glitchy results. Extra high took about an hour and a half and cost around $26, delivering the best outcome. Max took two and a half hours, cost about $50, and wasn’t meaningfully better, Anthropic’s own documentation even notes that max effort is prone to overthinking.
The practical takeaway is that model choice and effort level are two separate dials. Medium effort is a reasonable default for most everyday work, with higher effort reserved for genuinely hard problems. Combining a lighter model like Sonnet with the right effort setting often beats defaulting to a heavier model like Opus at max effort, both on cost and on actual output quality.
Frequently Asked Questions
Which Claude model should I use for everyday work?
Sonnet is the best default for most tasks, especially ones with clear instructions and a defined output. It’s generally faster and cheaper than Opus while producing comparable results on well-scoped work.
When should I use Opus instead of Sonnet?
Use Opus when the task is vague or open-ended, such as building a plan or strategy from scratch where the model needs to make structural judgment calls rather than follow explicit steps.
Is Fable better than Opus?
Fable is Anthropic’s most powerful model and is built for long, autonomous reasoning tasks, but Anthropic’s own release notes state that Opus performs near Fable level on most work. Fable also costs about two and a half times more than Opus, so it’s only worth it for tasks that genuinely need the extra capability.
Does the free Claude plan include Opus or Fable?
Run it on our own servers.
✓SOC 2 and HIPAA from day one.
✓Log every action.
✓Only the right people get in.
✓No. The free plan includes Sonnet and Haiku only. Opus and Fable require paid plans, and Fable access is further limited by weekly usage caps on Max and Ultra or billed through usage credits on Pro.
What is the /advisor command used for?
The /advisor command lets Claude check in with a stronger model, like Fable, at key points during a task. This gives you access to additional reasoning power without running the entire job through the more expensive model.
