Claude Haiku 5.5 for Copywriting: Cost, Rule Following and compared to Haiku 4.5

Claude Haiku 4.5 finished dead last in my writing benchmark. It used 29 em dashes in three blog posts and missed 8 of 15 Google Ads character limits. So when Anthropic released Haiku 5.5 on October 7, I wanted to see if any of that got fixed.
All of it did. Zero em dashes, all 15 limits. In short, Haiku 5.5 is now my pick for high-volume copy.
Is Claude Haiku 5.5 good at copywriting?
Yes. In my testing it scored 89.6 out of 100, 7th of 19. That’s the same score as Claude Sonnet 5.5, Draftly’s default writer. Haiku cost $0.0019 per copy pack. Sonnet cost $0.0301.
Here’s why it got that score.
The upgrade from Haiku 4.5 is huge
| Claude Haiku 4.5 | Claude Haiku 5.5 | |
|---|---|---|
| Copy score | 80.5 | 89.6 |
| Google Ads limits met | 7 of 15 | 15 of 15 |
| Blog score | 74.0 | 89.5 |
| Em dashes in 3 blog posts | 29 | 0 |
| Cost per blog post | $0.0100 | $0.0031 |
Haiku 4.5 is still last on blog posts. Haiku 5.5 placed 8th of 19, and it costs less.
What I tested
I added Haiku 5.5 to Draftly’s benchmark on October 20, with the same briefs, prompts and judges as the other 18 models.
Each model writes a copy pack for three different businesses:
- A roofer in Boise
- An accounts payable software company
- A dental practice in Phoenix
The pack covers the everyday copywriting jobs those businesses actually need:
- Five landing page headlines
- A hero section (headline, subhead and button)
- A Google search ad (three headlines and two descriptions)
- Three email subject lines with preview text
- Three benefit bullets
Every piece has a hard limit, like 30 characters for an ad headline or 15 words for a bullet. That’s 20 limits per pack and 60 across all three.
How the copy score works
The copy score is two numbers averaged together.
- Code score. Each pack starts at 100 and loses points for broken rules: 3 per missed limit, 5 per banned AI word, 5 per contrast phrase, 2 per em dash, 4 per brief fact it left out and 5 per near-duplicate headline.
- Judge score. Two judges from different companies, Gemini 3 Flash and GPT 5.4, rate each pack 1 to 10 on clarity, persuasion, voice and specificity. That’s 40 points, scaled to 100.
Here’s how Haiku 5.5 did on each pack:
| Pack | Code score | Judge score |
|---|---|---|
| Roofer in Boise | 100 | 79 |
| AP software | 100 | 81 |
| Dental practice | 100 | 78 |
| Average | 100 | 79 |
(100 + 79.2) ÷ 2 = 89.6.
The code found nothing to take off. It hit all 60 limits, used all 12 brief facts, and had no banned words, dashes or contrast phrasing. The judges gave it 8.5 for voice, 8.3 for clarity, 7.8 for specificity and 7.0 for persuasion.
What the judges liked
Its copy is warm and specific. Both judges picked this as the best line in the dental pack:
Put off the dentist? We’ll go at your pace.
That’s exactly the voice the brief asked for: warm, reassuring, plain. For the roofer:
A written estimate you can read before you decide.
And for the AP software company:
Three-way matching that catches mismatches before payment
What cost it points
Persuasion. At 7.0 out of 10, it’s Haiku’s lowest score, which is true of every model we tested. Some lines just state the category. Both judges picked “Family and cosmetic dentistry in Phoenix” as the weakest line in the dental pack. Every dentist in Phoenix could use it.
Is Claude Haiku 5.5 good at blog posts?
Yes, if you ask for a longer draft. It placed 8th of 19 at 89.5.
Each model wrote one blog post per business, with a 1,000 to 1,200 word target. The blog score averages two numbers:
- Formatting score. Each post starts at 100 and loses points for things like word count, dashes, contrast phrases, long paragraphs, a missing FAQ, keyword placement, meta title and description length, and a missing call to action.
- Voice score. Both judges run a 13-point human-voice audit, scored out of 100.
| Post | Words | Formatting | Voice |
|---|---|---|---|
| Roofer in Boise | 948 | 90 | 92.5 |
| AP software | 861 | 92 | 89 |
| Dental practice | 1,064 | 87 | 86.5 |
| Average | 958 | 89.7 | 89.3 |
(89.7 + 89.3) ÷ 2 = 89.5.
It ran short. Two of the three posts came in under 950 words and lost 5 points each for it. The roofing post also lost 5 for one contrast phrase. The AP post lost 3 for a missed keyword placement. The dental post lost 10 for two contrast phrases (“pressure rather than pain” is one) and 3 for keyword placement. That’s why it passed 16 of 18 SEO checks.
The writing itself is good. From the opening of the AP post:
When quantities, prices, and line items agree within your tolerances, the invoice moves to payment. When they don’t, an AP clerk opens three browser tabs and starts hunting.
That last sentence is something a person who has done AP would write.
How it compares to the leaders
| Model | Copy score | Judge score | Ads limits met | Cost per copy pack | Seconds per pack | Blog score | Cost per blog post |
|---|---|---|---|---|---|---|---|
| Claude Haiku 5.5 | 89.6 | 79 | 15 of 15 | $0.0019 | 15 | 89.5 | $0.0031 |
| Gemini 3.8 Flash | 90.8 | 82 | 15 of 15 | $0.0085 | 15 | 89.4 | $0.019 |
| Claude Opus 5.5 | 90.8 | 82 | 15 of 15 | $0.0532 | 24 | 93.8 | $0.099 |
| Kimi K3 | 90.0 | 80 | 15 of 15 | $0.0818 | 46 | 93.9 | $0.136 |
Costs are what the Vercel AI Gateway billed us, at $0.10 per million input tokens and $0.50 per million output. Blog posts took about 29 seconds each.
Who should use Claude Haiku 5.5
If you write in volume (ads, subject lines, benefit bullets, lots of variants), this is my pick. It’s within about a point of the leaders and costs a fraction of a cent.
For blog posts, Sonnet 5.5 is still better: 92.8 against 89.5, with a higher voice score, 95 against 89. But Sonnet costs about 16 times as much per post. For a lot of small businesses, Haiku 5.5 is good enough, as long as you ask for a longer draft.
If you dropped Haiku because of the dashes, the new one is worth another look. What would you use it for first?
A note on the method
Every model got the same system prompt (Draftly’s production writing rules) and the same three briefs, called through the Vercel AI Gateway. Three briefs per model is a small sample, so treat a gap of one or two points as a tie.
Full results, every model and the method: Best AI for writing: 19 models tested.
Claude Haiku 5.5 is available for AI copywriting in Draftly today. Sign up today with your free post to get started.
One practical SEO-writing tip, every other week
One tip per email. Unsubscribe anytime.
We'll email you once to confirm. By subscribing you agree to our Privacy Policy.
Keep reading
BenchmarkBest AI for writing: 19 models testedWhich models write the most human-sounding posts, and what each one costs.
Blog postHow to Use an SEO MCP Server With Claude and ChatGPT (Draftly MCP Announcement)An SEO MCP server gives Claude and ChatGPT live keyword data, Google results and competitor gaps. Here are Draftly's…
Blog postAEO and GEO for small business blogs in 2026How to get your small business blog cited by ChatGPT, Claude and Google AI Overviews, based on Google's own guidance…
