Claude models: Fable vs. Opus vs. Sonnet vs. Haiku

Anthropic's Claude family has four primary models: Haiku, Sonnet, Opus, and Fable. The version numbers are inconsistent between them; as I write this, Opus 5.5 just dropped, Fable 5.1 came out on September 1, 2026, Sonnet 5 was released on June 30, 2026, and Haiku is still on 4.5 and launched in October 2025—though a new model is due out soon. There's also Mythos, but the general public doesn't have access to that. 

All four Claude models are incredibly capable. Even the smallest model, Haiku, is better than the best frontier models from two years ago. This makes teasing out the differences between them a challenge. 

Benchmarks are one tool to help. Zapier's AutomationBench, for example, tests AI agents on end-to-end workflow execution using real tools across different business functions. Each task drops an agent into an isolated environment (CRM records, inbox threads, calendars, and more), and it's built on real patterns from 2B+ monthly tasks across 3.7M companies.

So if you're looking for a comparison of quality across these Claude models, here's how each model's most powerful version stacks up on AutomationBench.

But I wanted to put all these models through my own tests. To do so, I used the same five prompts for each.

• Categorize 25 customer support messages into Billing, Bug, Feature Request, Account Access, or Other with a confidence score. The model had to return a JSON array, and there were a few ambiguous messages. 

• An impossible calendar scheduling puzzle. This wasn't about getting the right answer, but seeing how each model handled an unsolvable problem. Earlier models would have just silently changed the constraints or claimed success when they clearly had failed. 

• Rewrite a product announcement to less than 150 words while avoiding jargon and cutting any unsupported or overblown claims. It then had to explain what it had cut and why. 

• Analyze an annual report, terms of service, and privacy policy for contradictions and disagreements. 

• Build a Kanban board in a single HTML file with no external libraries. It had to list any requirements it couldn't meet.

Of course, these tests aren't the only thing I'm basing this write-up on. I've been writing almost exclusively about AI for the past four years, I've spent hundreds of hours working with every Claude model, and my personal AI bill is at least $400/month. As well as some real-world results, I'll try to give you a practical framework for thinking about each model. 

To make things simple, I used the Claude app for the tests, but all the models are available through Claude Code and the API too. (For advanced uses, Claude Code is normally the most suitable tool.)

Now, let's get into it.

Table of contents:

• Claude Haiku vs. Sonnet vs. Opus vs. Fable at a glance

• Claude Haiku: Fast and cheap

• Claude Sonnet: The conversational default

• Claude Opus: For work and coding

• Claude Fable: Frontier coding

• Which Claude model is best?

• Add Claude to your workflows with Zapier

Claude Haiku vs. Sonnet vs. Opus vs. Fable at a glance

Claude Haiku: Fast and cheap

• Input price: $1/million tokens

• Output price: $5/million tokens

• Context window: 200K context, 64K max output

• Knowledge cutoff: February 2025

Claude Haiku is Anthropic's fastest and cheapest model, though that comes with a couple of caveats. It's due for an update very soon, so expect its performance to improve substantially. 

Haiku is best for simple and repeatable tasks, like categorizing messages, routing support tickets, and powering the real-time chat layer of a customer service app where speed matters more than depth. 

It correctly categorized all 25 customer messages, though it missed a few ambiguous cases where a second category was also appropriate. (Sonnet also missed the same cases.)

Haiku is a reasoning model. The catch is that it isn't very good at it compared to the other Claude models; it only supports "extended thinking" and doesn't allow you to set a reasoning effort. For the impossible puzzle, Haiku took almost two minutes to figure out that it was unsolvable—and even then, its reasoning was the weakest. Opus and Fable clocked the issues instantly, and Sonnet took less than 15 seconds. 

Perhaps the biggest problem was in the rewriting test: Haiku silently reworded a direct quote. This is the kind of thing that gets you in big trouble, and it also missed the word count.

Similarly, its document assessment and Kanban board were the weakest of the lot. It found the inconsistencies in the documents but didn't really understand the context of them, and search didn't work in its app. 

For tasks that suit Haiku, it's a strong model. But for tasks that are beyond its capabilities, it can fail without alerting you, and also take longer to get worse results than more advanced models. 

In my opinion, Haiku is only worth using through the API (or for an automation). It's not a model you want to work with on a daily basis; you want to assign it a very specific and limited task that it does automatically. 

Claude Sonnet: The conversational default

• Input price: $2/million tokens

• Output price: $10/million tokens

• Context window: 1M context, 128K max output

• Knowledge cutoff: January 2026

Claude Sonnet is a big step up from Haiku and a solid all-rounder. If you're on the free plan, it's the most advanced model you have access to. If you're on the $20/month Claude Pro plan or you want to conserve your token usage on a Max plan, it's a very good default model, especially through chat. 

I've also had good results getting Opus to delegate some coding tasks to Sonnet agents, but that's getting deep into the technical weeds. A new version of Sonnet is also due out soon.

Like Haiku, Sonnet correctly categorized all the customer support messages but missed two ambiguous cases. Its reasoning about the impossible puzzle was substantially quicker, and its rationale was much clearer. It also found the exact same proof that the puzzle was unsolvable as Opus.

In the rewriting and document analysis tasks, Sonnet showed its strengths. While it missed the word count by 17 words, it cut all the hyperbolic statements and left the direct quote intact. It also found all the contradictions in the three documents—and even found one bonus issue around the number of employees that the more advanced models missed. Its assessment of the issue priority fell short, but it was closer to Opus than Haiku.

It produced a working Kanban board and accurately assessed the limitations with its design, but it lacked the visual polish of Opus or Fable.

All told, Sonnet is a useful and flexible model. If budget isn't an issue, Opus and Fable generally outperform it—though by how much depends on the task. For API use, it's capable of more advanced linguistic reasoning than Haiku. I use it to generate SEO briefs from search data, and it consistently gives solid results.

Claude Opus: For work and coding

• Input price: $4/million tokens

• Output Price: $20/million tokens

• Context window: 1M context, 128K max output

• Knowledge cutoff: June 2026

Opus is only available to paid Claude subscribers (or through the API). For most people doing coding and development work, Opus at various different reasoning efforts will be the default model for essentially everything. 

Just this week, I've used Opus through chat to analyze my blood work, find a paint color match available in Ireland for British Racing Green, fact-check articles I'm writing, and figure out if there's an open-source alternative to Bird Buddy I can build. Could Sonnet have got good results for these? Maybe, but Opus aced them all first time and quickly enough that it fits with my workflow. 

In the categorization test, Opus correctly identified all the ambiguous cases. It instantly saw that the puzzle was unsolvable and proved it cleanly. It hit the word count for the rewriting task, and kept the quote intact. It found all the inconsistencies in the document tasks and correctly classified them by severity. Its Kanban board was fully functional and nicely styled.

Really, for most people, Opus will hit the balance between good enough at everything and not so expensive to run that it wipes out your usage instantly that it will be the model you want to use in most cases. Sometimes, you'll be token constrained or need Fable's power, but Opus is the Claude model I default to 95% of the time. (Opus 5.5 is also cheaper than Opus 5 and is more efficient at using tokens, further strengthening the argument to default to it.)

One thing to note: Opus 5 was infamous for its awful and long-winded Claudish writing style, but 5.5 seems to have fixed it. I suspect the issue was to do with optimizing Opus for reasoning and coding tasks.

Claude Fable: Frontier coding

• Input price: $10/million tokens

• Output price: $50/million tokens

• Context window: 1M context, 128K max output

• Knowledge cutoff: June 2026

Claude Fable is great at frontier coding and advanced long-running tasks, but it has some downsides too. For most tasks, Opus 5.5 matches it on benchmarks, and I'd say that's pretty true in my testing. It also doesn't take direction especially well. Like other experts, I've found it performs best when you give it a brief and let it do its own thing, rather than dictate every limitation or step it has to take. 

Where I've found Fable especially useful is in planning code. I've used it to create implementation plans, or review implementation plans that I can then delegate to other models to actually run. I'll also use it to do targeted code and feature audits. I feel it gives you a huge amount of Fable's advantages without the astronomical usage costs.

Part of the issue, I suspect, is that frontier models like Fable are trained to think hard about everything. Even simple tasks can turn into overwrought reasoning exercises, especially at higher reasoning levels. 

So how did it do on my tasks? About as well as Opus. It correctly categorized all the messages, including the ambiguous ones, and identified why the scheduling problem was unsolvable. In the rewriting task, it came in under the word count but made some small cuts to the direct quote—the kind of thing an editor would do, but also a little dicier than Opus's approach. In the document task, it found the same planted issues as Opus and prioritized them defensibly, though it made one unsubstantiated claim. Its Kanban board had the best design sensibilities, but both models still passed the test. 

Really, Fable is a great model, but if you don't know for sure you need it for a specific task or have unlimited budget, Opus is the better default. 

One quirk worth noting: you might end up using Opus anyway. Fable falls back to Opus when it detects a request that trips its safety guardrails. This includes some things around computer security, biology, and science and can be triggered by your prompt, the files you upload, or other details. It tells you when it happened, so it isn't hidden, but it's still worth understanding. 

Which Claude model is best?

Claude Haiku, Sonnet, Opus, and Fable all have legitimate use cases, though there's a huge amount of overlap in everyday capabilities. For lots of tasks, Fable and Opus don't significantly outperform Sonnet and Haiku. Of course, for more advanced tasks (especially coding), Fable and Opus offer substantial advantages—though the gap between them can be very narrow. 

Even in my testing, Haiku only failed the single most advanced task with a search bug (though it reported that it had succeeded)—the others passed everything. The difference wasn't in whether or not they gave an answer, it was in how accurately they followed instructions, what compromises they made, how they handled ambiguity, and how well they reported on what they'd done. For the most part, more advanced models provided more clarity around what they'd done—though Fable does march to the beat of its own drum more often than is ideal.

In most cases, it takes a bit of trial and error and some human judgment to work out which model is truly optimal. In general, though, Opus offers the most consistent balance between price and performance if you have a Claude subscription; if you don't, want to conserve tokens, or aren't worried about coding, Sonnet is a solid default. 

Add Claude to your workflows with Zapier

Whatever Claude model you use, you can use it in your automated workflows with Zapier. All four—Haiku, Sonnet, Opus, and Fable—work with AI by Zapier, which means you can bring Claude's brain into all your workflows, across your entire tech stack. Because each model excels at specific tasks, you can use different models for different steps in the same workflow, and swap between them at any point without rebuilding your workflows. 

Unless you bring your own key (BYOK), you're also not paying per token. Zapier bills by task—1, 3, or 5 tasks, depending on the model—so costs stay predictable no matter how much the model is doing under the hood.

Or, if you'd rather stay in the Claude app, Zapier MCP gives it secure access to 9,000+ apps, so you can take action straight from the chat window.

Related reading:

• What is Claude Code? The AI coding tool anyone can use

• Claude API: How to get a key and use the API

• Claude for small business: What it is and how to use it