Claude Haiku vs Sonnet vs Opus: Which Model Should You Use? (2026)

Anthropic offers three families of Claude models in 2026: Haiku, Sonnet, and Opus. The names suggest increasing sophistication (and price), and that’s exactly what they are. But the real question — which one should you use for your workflow — isn’t answered by a benchmark chart.

This post is the decision framework I use when I’m picking a model for a real Make.com scenario or API project. No speed tests, no contrived benchmarks. Just: when each one earns its cost.

Quick Overview (April 2026)

Model	Speed	Cost (input/output per 1M tokens)	When to Use
Haiku 4.5	Fastest	~$1 / ~$5	High volume, simple tasks
Sonnet 4.6	Balanced	~$3 / ~$15	The default for most work
Opus 4.7	Slowest	~$15 / ~$75	Hard reasoning, long analysis

(Prices change — verify current pricing at anthropic.com/pricing.)

Haiku is about 15x cheaper than Opus on output. That gap matters more than any capability gap for 90% of automation work.

When to Use Haiku 4.5

Haiku is the right choice when:

• The task is pattern-based: classification, extraction, summarisation, translation
• You’re processing high volume (hundreds or thousands of calls per day)
• Response time matters (Haiku is fast)
• The input and output are both relatively short

Real examples from my workflows:

• Classifying incoming emails as “sales”, “support”, or “spam”
• Extracting booking details from customer messages
• Generating short product descriptions from structured data
• Summarising customer reviews into one sentence
• Translating between languages for tourism content

All of these run on Haiku happily. Moving them to Sonnet would triple the cost without meaningfully improving output.

When Haiku falls short: multi-step reasoning (e.g. “read this contract, identify the three clauses that conflict with Australian Consumer Law”), nuanced tone (e.g. writing a sensitive customer apology), or long-form content.

When to Use Sonnet 4.6

Sonnet is Anthropic’s default recommendation for good reason — it’s the sweet spot between capability and cost for most real work.

Sonnet is the right choice when:

• The task requires reasoning, not just pattern-matching
• You need longer, more structured output
• Quality of writing matters (customer-facing content, marketing)
• Volume is moderate (tens to hundreds of calls per day)

Real examples:

• Drafting customer email replies that need the right tone
• Writing social media posts that match a brand voice
• Analysing a document and producing a structured report
• Generating code that needs to actually work
• Multi-turn conversations where Claude needs to remember context

When Sonnet falls short: very long documents (50k+ tokens), highly specialised domain reasoning (legal analysis, advanced maths), or tasks where Opus’s deeper reasoning is genuinely needed. But honestly — that’s less often than people assume.

When to Use Opus 4.7

Opus is expensive enough that using it for everything is a billing mistake. But there are tasks where nothing else comes close.

Opus is the right choice when:

• The task requires reasoning across a long document
• You need the best available analysis, not just a good one
• The work is one-off or low-volume (research reports, strategic analysis)
• The output is high-value enough to justify the cost

Real examples:

• Analysing a 100-page contract and identifying risks
• Strategic research: “Here’s my business and market — what would you prioritise?”
• Code architecture decisions across multiple files
• Synthesising a final report from 20 research sources
• Complex problem-solving where you can’t afford to be wrong

When Opus is overkill: pretty much everything else. A good heuristic: if the task is something you’d feel fine delegating to a capable junior, Sonnet is enough. If you’d only trust a senior expert, Opus might be worth it.

The Practical Decision Tree

Here’s how I actually decide in practice:

1. Is the task simple pattern-matching? → Haiku. Always start here.
2. Is Haiku’s output quality insufficient after prompt refinement? → Sonnet.
3. Does the task require deep reasoning over long content? → Opus.
4. Is the task mission-critical and will be done once? → Opus.
5. Is the task high-volume? → Haiku, with a good prompt, even if quality is 90% of Sonnet.

The key move is always starting with the cheaper model and only upgrading when the cheaper one demonstrably can’t do the job. The reverse approach (start with Opus, downgrade when you realise you’re overpaying) costs real money while you figure it out.

Model Routing in Practice

On production workflows, I often use two models in the same scenario:

• Haiku for classification: “Is this email a support ticket?”
• Sonnet for the reply: if yes, draft a response

This routes the expensive model only to cases that need it. For a workflow that handles 500 emails a week where 60% are spam, you’ve cut Sonnet calls by 60% — real savings.

Make.com makes this trivial: a Router module after the first Claude call, with filters based on Haiku’s output.

Don’t Fall for Benchmark Theatre

A benchmark that shows “Opus scores 92% vs Sonnet 88%” on some task doesn’t mean Opus is the right choice for your workflow. It means Opus is 4% better on that specific benchmark.

For most real work:

• Prompt quality matters more than model choice. A well-prompted Haiku often beats a badly-prompted Sonnet.
• Example-based prompting (giving Claude 2-3 good examples of the output you want) usually eliminates the need to upgrade models.
• Task decomposition (breaking a hard task into several easier ones) lets you use cheaper models throughout.

Anthropic’s own guidance is the same: try Haiku first, move up only when needed.

Switching Models in Make.com

In Make.com, changing the Claude model is a single dropdown — Model field in the module settings. You can A/B test in minutes:

1. Duplicate your scenario
2. Change the model on the duplicate
3. Run both against 10-20 test inputs
4. Compare outputs side by side

If Sonnet’s output is noticeably better for your case, keep it. If it’s indistinguishable from Haiku’s, save the money.

Next Steps

If you want the full decision framework — prompt patterns, model routing, A/B testing templates — that’s core material in the Implementation Blueprint ($29). Covers this in detail with real scenarios.

For a quick hands-on intro, the free Quick Start uses Haiku so you can feel how fast and cheap it is for simple tasks.

---

Last updated: 20 April 2026. Model lineup current as of this date — I update when Anthropic releases new models.