Claude Haiku 5.5 Costs One Tenth of Haiku 4.5. Here's the Margin Math for Every AI Feature You Decided Was Too Expensive.
Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Haiku 4.5 cost $1 and $5. That is a 10x cut on both sides, with a 1 million token context window on top. If you shelved an AI feature because the per-user cost didn't fit your pricing, it's worth redoing the math.
The numbers, from the pricing page
Anthropic's pricing page lists Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts of 100K tokens or fewer, with cache reads at $0.01. Past 100K tokens the rates rise to $0.50 input and $2.50 output. Sonnet 5.5 sits at $2 input and $10 output, and Haiku 4.5 is listed as legacy at $1 and $5. The page also advertises 50% off for batch processing, though it doesn't spell out per-model batch prices.
Coverage of the launch adds that the context window is 1M tokens (up from 200K), maximum output is 128K, and thinking is adaptive by default with an effort parameter that defaults to medium. I couldn't find a dedicated Anthropic announcement page with benchmarks, so I'm not going to tell you how Haiku 5.5 scores. Test it on your own prompts.
A worked example with stated assumptions
Say you run a support-reply helper inside a small SaaS. Assumptions: 1,000 requests a day, 3,000 input tokens and 500 output tokens each, no caching, no batch discount, 30 days a month.
That's 3 million input tokens and 0.5 million output tokens a day.
- Haiku 4.5: $3.00 input plus $2.50 output is $5.50 a day, or $165 a month.
- Haiku 5.5: $0.30 plus $0.25 is $0.55 a day, or $16.50 a month.
- Sonnet 5.5: $6.00 plus $5.00 is $11 a day, or $330 a month.
The same workload moved from $165 to $16.50. Per request, Haiku 5.5 costs about 0.055 cents.
What this unlocks
Features that were too expensive when you multiplied by every user, every page view or every background job:
- Run it on everything, not on request. Summarize every incoming ticket, tag every uploaded document, classify every form submission, instead of only when someone clicks a button.
- Free tiers that include AI. If a heavy free user costs you a few cents a month, you can afford the free tier that converts.
- Long documents. A 1M context window priced at $0.50 per million input tokens above 100K means stuffing a big document in the prompt costs real but bearable money.
- Retries and verification passes. Having a second call check the first is cheap enough to do by default.
What it doesn't change
Cheap tokens don't fix a bad feature. They don't fix latency either, or accuracy on hard reasoning. If your feature needs judgment (contract review, code changes that run unsupervised), Sonnet 5.5 at $2/$10 may still be the right call, and the cost difference is small compared with a wrong answer. A $330 month for the worked example above is still a rounding error next to one churned customer.
It also doesn't settle pricing strategy. When your costs fall 10x, competitors' do too. Anyone selling a thin wrapper around the same model can undercut you, and the margin you just gained is the margin they'll compete away. Defensible products sit on your data, your workflow and your distribution, not on the token price.
Where I could be wrong
Vendor prices change, and a 10x cut could come with a quality gap that shows up only on your inputs. Adaptive thinking also means output token counts can vary, and reasoning tokens bill as output. My example assumes 500 output tokens; if the model thinks more on your prompts, your real bill will land higher. Run a day of real traffic through it before you rely on my numbers.
What I'd actually do
I'd pick the one feature I cut for cost, wire it to Haiku 5.5 on a flag, and log tokens per request for a week. If the quality holds, I'd ship it, leave Sonnet 5.5 for the 5% of requests that fail a cheap quality check, and keep prices where they are for a quarter so the savings turn into margin instead of a price war.
Author
Lukas
@lukcombinator