Claude Haiku vs Sonnet vs Opus: Which Claude Model Should You Use?
Anthropic now ships four current Claude models, not three. Most guides still compare only Haiku, Sonnet and Opus, and they quote prices that changed months ago.
That matters because the wrong choice is expensive. On the same workload, the cheapest current model costs 100 times less per token than the most capable one. Pick too small and the output isn't good enough. Pick too big and you burn budget on tasks a smaller model would handle.
This guide covers the current lineup (Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1), what each costs, and a simple way to decide. Every number comes from Anthropic's official docs, linked at the end.
🎯 Quick Answer (30-Second Read)
- Start with Claude Opus 5.5 for most work. Anthropic's own docs recommend it as the default for most workloads.
- Use Claude Sonnet 5.5 when you need speed and lower cost with strong quality. It's half the price of Opus.
- Use Claude Haiku 5.5 for high-volume, latency-sensitive tasks such as classification, extraction and routing. It's the fastest and cheapest.
- Use Claude Fable 5.1 only for the hardest reasoning and long-running agent work, when Opus at higher effort still falls short.
- All four have a 1M-token context window, 128K max output, vision and tool use.
- Cut costs with the Batch API (50% off) and prompt caching before you downgrade the model.
The Current Claude Lineup at a Glance
| Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | Fable 5.1 | |
|---|---|---|---|---|
| Best for | Classification, extraction, routing | Speed + intelligence balance | Agentic coding, knowledge work | Hardest reasoning, long-horizon agents |
| Speed | Fastest | Fast | Moderate | Slower |
| Input price | $0.10 / MTok* | $2 / MTok | $4 / MTok | $10 / MTok |
| Output price | $0.50 / MTok* | $10 / MTok | $20 / MTok | $50 / MTok |
| Context window | 1M tokens | 1M tokens | 1M tokens | 1M tokens |
| Max output | 128K | 128K | 128K | 128K |
| Default effort | medium | high | medium | high |
| API model ID | claude-haiku-5-5 |
claude-sonnet-5-5 |
claude-opus-5-5 |
claude-fable-5-1 |
MTok = one million tokens. *Haiku 5.5 pricing applies to prompts up to 100,000 tokens. Longer prompts cost $0.50 input / $2.50 output per MTok.
The names hint at size: a haiku is a short poem, a sonnet a longer one, and an opus a major work. Fable is the newest tier and sits above Opus.
How the Four Models Differ
classify · extract · route} B -->|Yes| H[Claude Haiku 5.5
fastest · cheapest] B -->|No| C{Do you need speed and
lower cost more than max quality?} C -->|Yes| S[Claude Sonnet 5.5
speed + intelligence] C -->|No| D{Does Opus at higher effort
still fall short?} D -->|No| O[Claude Opus 5.5
recommended default] D -->|Yes| F[Claude Fable 5.1
hardest reasoning] style A fill:#0f172a,color:#ffffff,stroke:#334155 style B fill:#78350f,color:#ffffff,stroke:#f59e0b style C fill:#78350f,color:#ffffff,stroke:#f59e0b style D fill:#78350f,color:#ffffff,stroke:#f59e0b style H fill:#166534,color:#ffffff,stroke:#16a34a style S fill:#1e3a5f,color:#ffffff,stroke:#3b82f6 style O fill:#312e81,color:#ffffff,stroke:#6366f1 style F fill:#7f1d1d,color:#ffffff,stroke:#ef4444
Claude Haiku 5.5: when volume and speed matter
Haiku is built for tasks that are well defined and repeated thousands of times: tagging support tickets, extracting fields from documents, routing requests, moderating content.
It's the fastest model in the lineup and by far the cheapest. One thing to watch is that pricing rises fivefold for prompts over 100,000 tokens, so keep Haiku prompts short where you can.
Claude Sonnet 5.5: the fast all-rounder
Anthropic describes Sonnet as "the best combination of speed and intelligence". At half the price of Opus, it suits everyday coding help, chat products, summarisation and agents that need to respond quickly.
Claude Opus 5.5: the recommended default
Anthropic's docs say to start with Opus 5.5 for most workloads. It's aimed at long-running agentic coding and knowledge work: multi-step tasks, large codebases and analysis where quality matters more than speed.
Thinking is always on, and its default effort is medium. Raise effort when a task is genuinely hard, and lower it when you want faster, cheaper answers.
Claude Fable 5.1: for the hardest problems
Fable is Anthropic's most capable widely released model, for demanding reasoning and long-horizon agent work. It's the slowest and the most expensive, at 2.5 times the price of Opus.
The official guidance is to use it when your evaluations on Opus 5.5 at higher effort still fall short. In practice, most teams should prove they need Fable before paying for it.
What It Actually Costs: A Worked Example
Take a support tool that processes 1,000 tickets a day. Each ticket sends about 2,000 input tokens and gets about 300 output tokens back. That's 2 million input and 0.3 million output tokens a day.
| Model | Daily cost | Monthly cost (30 days) |
|---|---|---|
| Haiku 5.5 | (2 × $0.10) + (0.3 × $0.50) = $0.35 | ~$10.50 |
| Sonnet 5.5 | (2 × $2) + (0.3 × $10) = $7.00 | ~$210 |
| Opus 5.5 | (2 × $4) + (0.3 × $20) = $14.00 | ~$420 |
| Fable 5.1 | (2 × $10) + (0.3 × $50) = $35.00 | ~$1,050 |
Two caveats keep this honest:
- Thinking tokens count as output. Models that think before answering produce extra output tokens, so real bills run higher than this simple maths. Lower effort means less thinking.
- This is list price. The discounts below can cut it sharply.
Try it yourself
Use this formula for your own workload:
Cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)
Run it for your two most likely models. If the cheaper one passes your quality checks, the savings compound every day.
Effort: The Setting Most Developers Ignore
All current models support an effort setting: low, medium, high, xhigh and max. It controls how much the model thinks before answering.
- Low suits chat, classification and simple extraction: faster and cheaper.
- Medium / high suits most coding and analysis.
- xhigh / max is for hard problems where correctness matters more than cost.
Before you switch to a cheaper model, try the stronger model at lower effort. It's often the better trade.
Here's a minimal Python example that sets the model and effort with the official Anthropic SDK:
import anthropic
client = anthropic.Anthropic() # reads ANTHROPIC_API_KEY
response = client.messages.create(
model="claude-sonnet-5-5",
max_tokens=1024,
output_config={"effort": "low"},
messages=[{"role": "user", "content": "Classify this ticket as billing, bug or feature: 'I was charged twice this month.'"}],
)
# Thinking blocks can come before the answer, so pick the text block
print(next(block.text for block in response.content if block.type == "text"))Cut Costs Before You Downgrade
| Lever | Saving | Best for |
|---|---|---|
| Batch API | 50% off input and output | Jobs that can wait (reports, bulk tagging, evaluations) |
| Prompt caching | Cache reads cost 5% of input price on Opus 5.5 and Sonnet 5.5, 10% on Haiku 5.5, 2.5% on Fable 5.1 | Long system prompts or documents reused across requests |
| Lower effort | Fewer thinking tokens | Simple or well-defined tasks |
| Shorter prompts | Fewer input tokens | Everything, and especially Haiku (prices rise above 100K tokens) |
With batch, Opus 5.5 drops to $2 input / $10 output per MTok, the same as Sonnet 5.5's normal price.
Limitations and Trade-offs
- Slower models cost you latency, not just money. Fable and Opus are noticeably slower than Haiku and Sonnet, which matters in live chat.
- Haiku's long-context pricing is a trap. Above 100,000 input tokens, its price per token rises fivefold.
- Older models are still available. Opus 4.x, Sonnet 4.6 and Haiku 4.5 remain listed as legacy models, but new projects should start on the current lineup.
- Prices and models change. Check Anthropic's pricing page before budgeting a production system.
Example Scenario
This is an illustrative example, not a real case study.
A two-person team builds an AI support desk. They route every incoming ticket through Haiku 5.5 to classify it, send standard billing questions to Sonnet 5.5 for a fast drafted reply, and escalate complex technical issues to Opus 5.5. Classification runs overnight through the Batch API at half price. The expensive model only sees the small share of tickets that need it.
Frequently Asked Questions
Which Claude model is best for coding?
Anthropic recommends Opus 5.5 for long-running agentic coding and most workloads. Sonnet 5.5 is a good choice when you want faster, cheaper help for everyday coding. Fable 5.1 is worth testing only if Opus at higher effort still fails your hardest tasks.
What is the cheapest Claude model?
Claude Haiku 5.5, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Longer prompts cost $0.50 input and $2.50 output per million tokens. With the Batch API, those prices drop by half.
What is Claude Fable and how is it different from Opus?
Fable 5.1 is the tier above Opus, built for demanding reasoning and long-horizon agent work. It costs 2.5 times as much as Opus 5.5 ($10 input / $50 output per million tokens) and runs slower. Anthropic suggests it when Opus 5.5 at higher effort still falls short.
Do all Claude models have the same context window?
Yes. Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1 all support a 1-million-token context window and up to 128,000 output tokens per request. All four also support vision and tool use.
Is Sonnet or Opus better value?
It depends on your quality bar. Sonnet 5.5 costs half as much as Opus 5.5 and responds faster. Opus is the better default for complex multi-step work. Test both on 20–30 real examples from your app and pick the cheaper one that passes.
Conclusion
The Claude lineup is no longer just Haiku, Sonnet and Opus. Today it's Haiku 5.5 for volume, Sonnet 5.5 for speed, Opus 5.5 as the recommended default, and Fable 5.1 for the hardest problems.
Start with Opus 5.5, measure, then move down a tier only where quality holds. Use batch processing, prompt caching and lower effort before downgrading. Those levers often save more than switching models.
Sources
- Claude models overview, Anthropic docs (model IDs, context windows, descriptions, default effort)
- Claude pricing, Anthropic docs (token prices, batch discount, cache pricing, Haiku long-context pricing)
- Choosing a model, Anthropic docs
Prices and model details verified on 14 October 2026.
Related reads: Claude Code CLI Tutorial · Complete Claude Guide: Features, Skills, Projects · How to Create a CLAUDE.md File · Best AI Coding Tools for Developers · RAG Explained
Building with Claude? I write about AI dev tools and how real products are built. Follow along on LinkedIn or X, or invite me to speak.