Claude Haiku vs Sonnet vs Opus: I Priced One Marketing Job on All Three. The Sticker Gap Is 40x.

On October 7, Anthropic released Claude Haiku 5.5 and called it “the cheapest, fastest, and most capable small model we’ve ever released.” If you are weighing Claude Haiku vs Sonnet, the first number is the price: $0.10 per million input tokens, against $2.00 for Sonnet 5.5 and $4.00 for Opus 5.5.

Cheapest, fastest and most capable is three superlatives in one sentence, so read the pricing page for the footnotes. Anthropic’s own benchmark table for the new model mixes coding, computer-use and knowledge-work tests, and none of them is a marketing task.

Marketing work is different: sort this, summarize that, pull the facts out, draft the first version, judge the final one. So I priced one marketing job on all three models.

The sticker gap is 40x between Haiku and Opus. The bill is smaller or larger depending on prompt length, retries and how often you reuse the same context. Pick the cheapest model that still passes your own test.

I read Anthropic’s launch post, its pricing page and its models overview, plus the Artificial Analysis page on Haiku 5.5. I checked every quoted price and benchmark against its source on October 14, 2026.

The job costs are my own arithmetic. Prices are API prices in US dollars per million tokens (MTok).

Key Takeaways

• List prices per MTok: Haiku 5.5 $0.10 in and $0.50 out, Sonnet 5.5 $2 and $10, Opus 5.5 $4 and $20.
• The ratios: 20x between Haiku and Sonnet, 40x between Haiku and Opus, on input and output alike.
• Long prompts shrink the gap. Above 100,000 tokens Haiku 5.5 costs $0.50 and $2.50, so the gap to Sonnet falls to 4x. Sonnet and Opus have no such surcharge.
• Caching and batching can move the bill as much as the model choice does. A cached Sonnet run can cost the same as an uncached Haiku run.
• Anthropic positions Haiku 5.5 for narrow, repetitive work: summaries, classification, extraction, routing.
• A quick test beats a leaderboard. Run 20 of your own items through all three, score them blind, and keep the cheapest model that clears your pass mark.

What Anthropic Sells, and at What Price

Anthropic’s current lineup has four models. Here are the list prices from its models overview and pricing page, with Anthropic’s one-line description from the overview. Prices as of October 14, 2026.

All four list a 1M-token context window and 128K tokens of maximum output. Output costs five times input on every row.

Anthropic’s words for the small model are narrow. Haiku 5.5 “reliably handles quick and repetitive workloads (like summaries, compactions, database queries, and classification requests).”

Anthropic also says Sonnet 5.5 and Opus 5.5 “remain better choices for complex agentic coding tasks.” Its docs suggest starting with Opus 5.5 for most workloads. My split below is a cost-driven marketing choice, not Anthropic’s advice.

What Changed on October 7

Four things changed on launch day.

• The price. Haiku 5.5 charges $0.10 input and $0.50 output for prompts up to 100,000 tokens. Haiku 4.5 charged $1 and $5, so the sticker cut is 90%.
• The dial. Anthropic says Haiku 5.5 is “our first Haiku-class model to come with an adjustable effort setting.” You can tell the small model how hard to think.
• The surcharge. Anthropic prices Haiku 5.5 by prompt length: a request whose prompt is over 100,000 tokens pays $0.50 and $2.50. Anthropic says about 90% of requests to the previous Haiku stayed under that line.
• A change to Sonnet. Anthropic halved the Sonnet 5.5 cache-read price to $0.10 per MTok and says that lowers the cost of most agentic tasks by about 20%. Cache reads on Haiku 5.5 cost $0.01.

The Sticker Price Is Not the Bill

Two effects eat into the 20x gap. A third explains why Anthropic’s own saving figure is not 90%.

Long prompts. Feed Haiku 5.5 a 150,000-token archive of brand documents and it bills $0.50 per MTok for input, not $0.10. Sonnet 5.5 bills $2 for the same tokens, so the gap is 4x.

Verbosity and retries. Artificial Analysis scores Haiku 5.5 at 43 on its Intelligence Index. It reports that the model generated 440 million tokens to complete the index, against a median of 95 million for comparable models, and calls that verbose.

That comparison is with the median of the models Artificial Analysis lists, not with Sonnet. It still shows how a cheap model can spend more tokens, and every extra token and every retry is billed.

The tokenizer, which does not change the Haiku-to-Sonnet gap. Anthropic says Haiku 5.5 “has an updated tokenizer (similar to Sonnet 5.5’s and Opus 5.5’s), which means it uses slightly more tokens per task.” Sonnet and Opus use a similar one, so it does not narrow the gap between the three. It is one of two reasons why Anthropic’s average saving against the old Haiku is “around 75% less to run,” not 90%: the other is the smaller 50% cut on prompts over 100,000 tokens.

The $0.21 that Artificial Analysis lists is the average cost of one task in its own test suite, run at maximum effort. A marketing task costs something else.

If you need US-only inference, Anthropic’s pricing page adds a 1.1x multiplier, and on Haiku 5.5 it also applies to the higher over-100,000-token prices.

What the Benchmarks Say, and Who Ran Them

Every number in this table comes from Anthropic’s own launch post of October 7. They are the vendor’s chosen tests, so treat them as a claim, not a verdict.

Haiku 5.5 is far better than Haiku 4.5. Sonnet 5.5 still leads on every row.

On the two knowledge-work scores, the gap to Sonnet is 220 and 246 points. Haiku 5.5 gained 885 and 964 points over Haiku 4.5 on the same scores, so most of the distance is already closed. On agentic coding the gap is the widest of the percentage rows: 70.6% against 39.2%.

Anthropic’s table also lists OpenAI’s cost-efficient model, GPT-6 Luna, and Haiku 5.5 scores higher on every row where Luna has a figure. These are Anthropic’s chosen tests. I did not look for an independent head-to-head, so I make no claim about it.

Anthropic’s post also quotes HubSpot (92.8% on its CRM simulation suite) and AlphaSense. Both are vendor-chosen quotes, and neither is a marketing test.

One Job, Priced Three Ways

Take a job every marketing team has: sort 10,000 inbound items, such as reviews, form submissions or support emails, into categories and write a one-line summary for each.

These are my assumptions, illustrative and not measured: 300 tokens in and 50 tokens out per item. That is 3 MTok of input and 0.5 MTok of output in total, at list prices.

That is exactly the 20x and 40x of the price list. List price times volume shows nothing the price list does not.

Now add the effects from the previous section, again as my assumptions, and keep Sonnet unchanged. Say Haiku writes twice the output tokens at the same effort, so its output cost doubles to $0.50 and the total is $0.80. Add a retry on 5% of items and the total is about $0.84.

Against $11 for Sonnet that is a gap of about 13x, not 20x. The sticker gap shrinks. It does not vanish.

The Batch API takes 50% off input and output for work that can wait. Anthropic describes it as “asynchronous processing of large volumes of requests.” On the first table, the Haiku run in batch costs about $0.28.

Cost scales with your volume and your habits. Run this daily, add a second pass, let an agent retry on errors, and a model 40 times pricier than you needed becomes a line item someone asks about.

Caching: The Discount You Have to Switch On

Marketing prompts repeat. The same brand guide, the same tone rules and the same examples ride along with every item. That repeated block is what prompt caching is for.

These numbers are my illustration. Add a 4,000-token brand guide to every request of the sorting job. Across 10,000 items that is 40 MTok of repeated input.

At list prices without caching it costs $4 on Haiku 5.5, $80 on Sonnet 5.5 and $160 on Opus 5.5. With caching you pay read prices: $0.01 per MTok on Haiku 5.5, $0.10 on Sonnet 5.5 and $0.20 on Opus 5.5. The same 40 MTok then costs $0.40, $4 and $8.

A cache write costs more than a normal read. Anthropic’s pricing page lists 1.25 times the input price for a five-minute cache write, which is about a cent on Sonnet for a 4,000-token guide each time the cache has to be rebuilt. Check the caching documentation for any minimum prompt length before you rely on a short block.

Treat $0.40, $4 and $8 as the best case. On Opus 5.5 and Sonnet 5.5, Anthropic says a cache hit costs 5% of the standard input price.

A cached Sonnet run costs $4, and so does an uncached Haiku run. Both levers are worth a factor of 20, and you can pull both.

What a Cheap Model Makes Affordable

A price this low changes which jobs are worth automating at all. Take a search terms report with 100,000 queries and tag each one by intent: informational, commercial, transactional, navigational or junk.

Assume 100 tokens in per query, which is tight for five categories with examples, so read it as a floor. Add 10 tokens out. That is 10 MTok in and 1 MTok out, and at list prices Haiku 5.5 costs $1.50, Sonnet 5.5 costs $30 and Opus 5.5 costs $60.

At $1.50 nobody needs a budget. At $60 someone may want one, and a budget conversation can end a cheap experiment.

The real saving is that you rerun the job. Categories change, a campaign ends, a new product launches. When a full pass costs the price of a coffee, you rerun it every month and the report stays honest.

Claude Haiku vs Sonnet: Where Each Model Belongs in a Marketing Team

This split is my judgment, not a test result. It follows the job descriptions above and the benchmark table, and it stretches both where Anthropic’s pages say nothing.

Haiku 5.5: the sorting and reading desk. Jobs where a wrong answer is cheap to catch:

• tagging and routing leads
• classifying reviews by topic and sentiment
• pulling names, dates and amounts out of emails and PDFs
• first-pass summaries of long threads
• the reading stage of a research pipeline, where the agent only extracts quotes

Sonnet 5.5: the daily driver. Drafting copy that a human will edit, analysis of a campaign export, briefs, outlines, translations and agentic work where a tool call has to succeed. It costs $2 and $10 and leads Haiku on Anthropic’s agentic tests. If you pay for one model, start here.

Opus 5.5: the one to try on review. Anthropic’s launch post groups Sonnet 5.5 and Opus 5.5 together as the better choices for complex agentic coding, so this role is a guess about where the extra $2 and $10 might pay off: strategy reviews, the skeptic who attacks a draft, hard multi-step builds, and decisions where a mistake costs real money. Keep it only if your test shows it catches more.

At $4 and $20 you do not run Opus on 10,000 items. You run it on the ten that matter.

My reading agents run at low effort and my reviewer runs at high effort. I have not tested which model should do which job.

A task goes down a tier when a human checks it anyway, and up a tier when nobody does. A summary you will read before it goes out can come from the cheap model. A sentence that goes straight to a customer deserves the better one.

Nobody reads 10,000 tagged reviews one by one, so check a sample. Pull 50 outputs a week and read them. Bulk sorting stays on Haiku only while the sample stays clean, and meta descriptions that publish unread belong on Sonnet.

Watch the other side of that trade. The most expensive token is the one a human has to fix, and the money you saved on tokens returns as review time.

If you are building the setup behind this, my marketing tech stack shows where the model layer sits, and the AI marketing playbook covers what to automate first.

The Second Dial: Effort

Model choice is one dial. Effort is the other. Haiku 5.5 now has it, so the small model can think less on easy work and more on hard work.

Effort and model size both raise quality and the bill. They are separate dials, and you should set both on purpose.

Artificial Analysis evaluated Haiku 5.5 at its maximum setting. At that setting the same page reports 332 seconds to first token against a median of 1.83 seconds, which presumably includes thinking time. I would not copy a benchmark’s settings into a live workflow.

In one of my sessions, effort was inherited. In my CLAUDE.md article I describe how ten research agents launched in parallel and took on the session’s maximum effort. I stopped it after six minutes: about 31 million tokens, about 17 dollars on the usage meter, zero finished dossiers.

Set effort on purpose for every automated job: low for sorting and extraction, medium for drafting, high for review. As I noted in that article, Anthropic’s Opus 5.5 guide says effort level names do not map to the same amount of thinking across models, so test the setting on each model instead of copying it.

A Quick Test: Pick the Cheapest Model That Passes

Leaderboards do not know your brand voice, your product names or your customers. Your own items do.

Here is the test I would run before moving any recurring job to a cheaper model. Expect about an hour the first time and about 30 minutes after.

1. Collect 20 real items across the range: five easy, ten typical, five ugly.
2. Run the same prompt on all three models, with the same instructions, the same examples and effort set on purpose.
3. Strip the model names and score blind: pass or fail, with one sentence on why.
4. Set your pass mark before you look. I would start at 18 of 20 for internal work and 20 of 20 for customer-facing text.
5. Take the cheapest model that clears the mark and multiply its cost per item by your real monthly volume.
6. Date the result and write down the settings. When a new model ships, rerun the same 20 items.

Twenty items is a smoke test, not statistics. If two models are close, run 50.

App or API: Which Price Applies to You

Everything above is API pricing. It applies when a script, an automation tool or an agent calls the model and you pay per token.

If you only chat in the Claude app, you pay a subscription instead. The model picker then changes quality, speed and how quickly you use up your plan’s limits, not a per-token bill. Check which models your plan offers.

In my article on using AI for marketing I put numbers on heavy professional use of frontier models: roughly 200 euros a month, and 400 to 600 for a hyper user. A basic subscription is about 20 euros a month, as I wrote in ChatGPT for marketing.

When a script or an automation tool calls the model through the API, you cross over to API pricing, and model choice starts to show on an invoice.

Frequently Asked Questions

Claude Haiku vs Sonnet: which should I use for marketing?

Use Haiku 5.5 for sorting, extraction and first-pass summaries, Sonnet 5.5 for drafting and analysis, and try Opus 5.5 on reviews and decisions where a mistake is expensive. Then run the quick test above on your own items and keep the cheapest model that passes.

Is Claude Haiku 5.5 good enough for writing?

Anthropic positions it for quick, repetitive, high-volume tasks and does not mention brand writing. Test it on your own copy with the blind method above. For text that goes straight to customers, start with Sonnet and move down only if the test says so.

How much cheaper is Haiku 5.5 than Sonnet 5.5?

At list price it is 20 times cheaper on input ($0.10 vs $2) and on output ($0.50 vs $10) for prompts up to 100,000 tokens. Above that line Haiku 5.5 costs $0.50 and $2.50, so the gap is 4x. Retries and long outputs can narrow it further, and Anthropic changes prices often, so check its pricing page before you budget.

My Verdict

Claude Haiku vs Sonnet is the wrong fight if you pick a winner. Pick the cheapest model that passes your own test, per task, and write down the date.

The card:

1. Sort and extract on Haiku, and read a sample of the output every week.
2. Draft and analyze on Sonnet.
3. Try Opus on review, and keep it only if your test shows it catches more.
4. Set effort on purpose, and never let an agent inherit it.
5. Retest when a model ships: twenty items, one new price.

Prices and benchmarks on this page are as of October 14, 2026. Anthropic changes both often, so check the dates before you rely on a number.

Running AI inside a marketing team and want a second pair of eyes on which model does which job, and what it costs? Get in touch.