Claude Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5: which to use for what

Anthropic released three new Claude models in just over two weeks: Opus 5.5 on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. If you use Claude for your business, you now face a menu where the most expensive option costs forty times the cheapest per word written. Pick the wrong one and you either overpay for simple jobs or get weak results on hard ones.

The good news is that the choice is simpler than the names suggest. Here is what each model is, what it costs, and how I decide which one gets which job.

The three models in one table

All three come from the same family, so they share a lot. Each can read up to 1 million tokens at once (a token is a chunk of text; Anthropic's docs say 1 million tokens is roughly 555,000 words on the current tokenizer), and each can write up to 128,000 tokens in one answer. All three read images as well as text, and all three can use tools.

What differs is speed, depth and price. Prices below are Anthropic's API prices per million tokens.

Two terms in that table need a word of explanation.

An agent is an AI that works through a task in many steps on its own, using tools such as a browser, files or code, instead of answering one message at a time. "Agentic work" is that kind of multi-step job.

Effort is a setting that controls how hard the model thinks before it answers. Higher effort gives better answers to hard problems and costs more time and tokens. Anthropic's guide says tuning effort is often a better lever than switching models, which is worth remembering before you jump to Opus.

Claude Opus 5.5: for the hardest, longest jobs

Opus is the one to reach for when a task is long, has many steps, and an error is expensive. Anthropic built it for agents that run for a long time: large coding projects, research that ends in a finished document, work across many files.

Anthropic's launch post says Opus 5.5 costs 40% less to run than Opus 5 and writes its output more than 30% faster. On Anthropic's own tests it scored 81.8% on OSWorld 2.1, a test of using a computer like a person does (clicking, typing, moving between apps). There is also a fast mode on the Claude API at $8 in and $40 out per million tokens, for up to 2.5 times the output speed.

For a marketer or business owner, Opus earns its price on jobs like:

• Turning a messy folder of sales notes, spreadsheets and call transcripts into a market report.
• Building or fixing a website or internal tool with a coding agent such as Claude Code.
• Long research tasks where it needs to read many sources and keep track of them.

Claude Sonnet 5.5: the everyday workhorse

Sonnet is half the price of Opus and Anthropic calls it "the best combination of speed and intelligence". Its suggested uses in Anthropic's model guide are code generation, data analysis, content creation, visual understanding and agent tool use.

The jump from the previous Sonnet is large. Anthropic reported Sonnet 5.5 at 70.6% on Terminal-Bench 4.0, a coding benchmark, against 10.3% for Sonnet 5, and output more than 30% faster at the same price as Sonnet 5. On that same benchmark Anthropic's table puts Sonnet 5.5 slightly ahead of Opus 5.5's 66.4%, which tells you how close the two now are on many tasks.

For most writing, planning and analysis, Sonnet is where I would start:

• Campaign plans, ad copy, emails, blog drafts and scripts.
• Reading a spreadsheet export and explaining what changed.
• Reviewing a landing page screenshot and suggesting fixes.

Claude Haiku 5.5: cheap enough to run on everything

Haiku is built for volume. At $0.10 in and $0.50 out per million tokens for prompts up to 100,000 tokens, it costs a tenth of the previous Haiku 4.5 ($1 and $5). Above 100,000 tokens in one prompt the price rises to $0.50 and $2.50, so keep Haiku's jobs short.

Cheap does not mean weak any more. On Anthropic's OSWorld 2.1 offline test, Haiku 5.5 scored 72.4% against 15.7% for Haiku 4.5. It is also the first Haiku with an adjustable effort setting.

Haiku is right for jobs you do hundreds or thousands of times:

• Tagging every incoming lead or support message by topic and urgency.
• Pulling names, dates and amounts out of forms or invoices.
• Writing short product descriptions from a spec sheet, in bulk.
• Acting as a helper inside a bigger agent, doing the simple sub-tasks while Opus or Sonnet handles the hard part.

A worked example: one campaign, three models

Say you are launching a new service and need a launch plan, ten ad variations, and replies sorted for the 2,000 enquiries you expect in the first month. This is how I would split it:

The Haiku job costs very little. If each enquiry plus instructions is about 800 tokens in and 50 tokens out, 2,000 enquiries is 1.6 million tokens in and 0.1 million out: 1.6 × $0.10 + 0.1 × $0.50 = $0.21 for the month. The same job on Opus would be 1.6 × $4 + 0.1 × $20 = $8.40. Neither number is big, but at a hundred times the volume the difference becomes real money.

Two ways to cut the bill on any of them

Anthropic's pricing page lists two discounts that apply to all three models:

• Batch processing. If you don't need the answer straight away, the Batch API halves the price of input and output. Overnight jobs like tagging yesterday's enquiries or rewriting a product catalogue fit well here.
• Prompt caching. If you send the same long instructions or documents again and again, the repeated part can be cached and read back at a fraction of the price. For Sonnet 5.5 a cache read is $0.10 per million tokens against $2 for fresh input; for Haiku 5.5 it is $0.01.

These matter more to developers using the API than to people in the chat app, where you pay a monthly plan.

My rule of thumb

Anthropic's model guide describes two ways to start: begin with Haiku and move up only when it falls short, or begin with Opus and move down once the task works. For business users I simplify that to one question per job:

1. Is it a simple, repeated task where volume and price matter? Haiku.
2. Is it everyday writing, analysis or coding? Sonnet.
3. Is it long, complicated, multi-step, or costly to get wrong? Opus.

When in doubt, start with Sonnet, check the result, and only move to Opus if it is not good enough. Then test Haiku on any task you repeat often, because the savings add up there.