Business SystemsGuide

Claude Haiku 5.5 Costs 90% Less. One Line in the Price List Decides Whether You Get That.

Most of the coverage stopped at the headline price. The interesting part is a threshold: a prompt that sat comfortably at 85,000 tokens on the old model can cross 100,000 on the new one without a single word changing.

Author

DueClix Engineering

Published

Reading time

3 min read

On 7 October Anthropic released Claude Haiku 5.5, its small, fast model. The price is what got shared: $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.5. Anthropic's own summary is more careful. It says the new model costs "around 75% less" to run on average, not 90%. The gap between those two numbers is where a business either saves a lot of money or not much.

The two details that change the bill

  1. The price depends on prompt length. A request whose prompt is up to 100,000 tokens pays $0.10 and $0.50. Over 100,000, the same request pays $0.50 and $2.50, five times more. Anthropic's pricing page says the whole prompt counts, including the cached part, and each request is priced on its own.
  2. The same text now counts as more tokens. Haiku 5.5 uses Anthropic's newer tokenizer, which the pricing page says produces approximately 30% more tokens for the same text. The exact increase depends on the content.

Put those together and something non-obvious happens. A prompt of about 77,000 tokens on Haiku 4.5 can measure around 100,000 on Haiku 5.5. Nothing about the document changed. It simply crossed the line.

Example one: a support assistant

Say a business answers 10,000 customer questions a month. Each request sends about 6,000 tokens (instructions plus the help articles it looked up) and gets back a 400-token answer.

Haiku 4.5Haiku 5.5
Tokens per request (in / out)6,000 / 400~7,800 / ~520
Input cost60M × $1 = $60.0078M × $0.10 = $7.80
Output cost4M × $5 = $20.005.2M × $0.50 = $2.60
Monthly total$80.00$10.40

That is 87% less, even after the tokenizer inflation. Short prompts in high volume are exactly what this model was priced for.

Example two: reading long documents

Now a business sends 1,000 long contracts a month, each about 85,000 tokens on the old tokenizer, and asks for a 2,000-token summary.

Haiku 4.5Haiku 5.5
Tokens per request (in / out)85,000 / 2,000~110,500 / ~2,600
Price tierStandardOver 100K
Input cost85M × $1 = $85.00110.5M × $0.50 = $55.25
Output cost2M × $5 = $10.002.6M × $2.50 = $6.50
Monthly total$95.00$61.75

Still cheaper, but only 35% less, not 90%. If the same work were split so each request stayed under 100,000 tokens, the input would cost around a fifth of that. Whether the document can be split without hurting the answer is the real question, and it is an engineering question, not a pricing one.

What to do with this

  • Find your largest prompts. If your system logs token counts, look at the top 5%. Anything above roughly 75,000 tokens on an older Claude model is at risk of crossing the line.
  • Send less, not more. A retrieval system that picks the right three passages is cheaper than one that pastes in the whole manual, and usually more accurate too.
  • Use batch for anything that can wait. Anthropic's batch prices for Haiku 5.5 are half the standard rate: $0.05 and $0.25 under 100,000 tokens.
  • Test quality, not only price. A cheaper model that gets 5% more answers wrong can cost more in support time than it saves in tokens.

We wrote about what running AI in production actually costs last month. The model price was never the largest line, and it just got smaller. The prompt you send is now what decides the bill.

Sources

Written by

DueClix EngineeringEngineering team

The team that designs and builds DueClix systems. We write about the parts of the work that are worth writing down.

Have a process worth improving?

Let's build the system behind it. Tell us what the process is and where it breaks — the first conversation is about constraints, not technology.

Start a projectFixed-price checks