Claude Haiku 4.5 → Claude Haiku 5.5: cost and quality on real tasks
Claude Haiku 5.5 is a cheaper model from Anthropic: on the six tasks below it costs about 87% less than Claude Haiku 4.5 per request at list prices. Cheaper models can be weaker on harder inputs, so the question is whether it still works first time often enough.
Price per million tokens
Standard list prices, checked 2026-11-12. Prices change often, so check the source before you rely on them.
| Model | Input | Cached input | Output | Source |
|---|---|---|---|---|
| Claude Haiku 4.5 (current)anthropic/claude-haiku-4.5 | $1.00 | $0.10 | $5.00 | Anthropic's pricing page |
| Claude Haiku 5.5 (new)anthropic/claude-haiku-5.5 | $0.10 | $0.01 | $0.50 | Anthropic's pricing page |
- Note on Claude Haiku 4.5: Older model still offered. Note Haiku 5.5 is 10x cheaper for prompts under 100K tokens.
- Note on Claude Haiku 5.5: Price applies to prompts up to 100K tokens; prompts over 100K cost $0.50 in / $0.05 cached / $2.50 out.
Cost per 1,000 results on six common tasks
What 1,000 requests cost at list price, for six task shapes from different kinds of app. The token counts are illustrative, not measured: your prompts may be longer or shorter, and the switch test below shows your real ones.
The Claude Haiku 5.5 column counts 30% more tokens for the same text, because Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens. The exact increase depends on your text.
| Task, and what goes in and out | Tokens in / out | Claude Haiku 4.5 | Claude Haiku 5.5 | Change | Break-even: Claude Haiku 5.5 costs no more per result that works if it works first time at least |
|---|---|---|---|---|---|
| Support bot answerInstructions, three help-centre passages and the customer's question in; a short answer out. | 3,000 / 300 | $4.50 | $0.58 | 87% less | 13% as often as Claude Haiku 4.5 |
| Invoice field extractionInstructions and one invoice's text in; about 15 fields as JSON out. | 2,500 / 250 | $3.75 | $0.49 | 87% less | 13% as often as Claude Haiku 4.5 |
| Lead scoringScoring rules and one lead's form answers in; a score and a one-line reason out. | 1,200 / 100 | $1.70 | $0.22 | 87% less | 13% as often as Claude Haiku 4.5 |
| Email reply draftInstructions and the email thread so far in; a reply of about 300 words out. | 1,500 / 400 | $3.50 | $0.46 | 87% less | 13% as often as Claude Haiku 4.5 |
| TranslationInstructions and about 500 words of text in; the same text in another language out. | 900 / 800 | $4.90 | $0.64 | 87% less | 13% as often as Claude Haiku 4.5 |
| Search re-rankingA search query and 20 result snippets in; the ids of the best five, in order, out. | 4,000 / 50 | $4.25 | $0.55 | 87% less | 13% as often as Claude Haiku 4.5 |
Real cost per result: when a switch really saves
Real cost per result is what you actually pay for each result that worked, counting the failed tries. A result that fails gets tried again, and you pay for both tries.
So the price per request is only half the answer. Claude Haiku 5.5 stays cheaper per result that works as long as it works first time at least 13% as often as Claude Haiku 4.5. That's the new cost divided by the current cost, shown for each task in the last column above.
With a gap this big, cost per result rarely decides the switch. What decides it is whether your users notice: answers that fail first time cost you trust and paying customers, not just tokens.
Example, with a made-up rate
Say 8 in 10 of Claude Haiku 4.5's support bot answers work first time. You pay $4.50 for 1,000 tries and get 800 that work, a real cost of $5.63 per 1,000 support bot answers that work. Claude Haiku 5.5 costs $0.58 for 1,000 tries, so it matches that real cost if 11% of its support bot answers work first time, and costs less if more do.
Quality: test it on your own requests
This page can't tell you whether Claude Haiku 5.5 handles your requests as well as Claude Haiku 4.5, and we don't publish quality scores. Quality depends on your prompts and your inputs: a model that answers short English questions well can still fail on long invoices, odd layouts or other languages.
The free model switch test runs your prompt and up to 20 of your real inputs on both models with your own key, through each provider's own API or OpenRouter. You see the answers side by side with the tokens and cost of each, and mark which ones worked.
Test both models on your promptsWhat to compare
- Format: does JSON still parse with the same field names, for extraction and scoring?
- Length and tone: are support answers and email drafts as short and polite as before?
- Refusals: does Claude Haiku 5.5 decline requests that Claude Haiku 4.5 handled?
- Hard inputs: long documents, other languages and messy formatting, not just the easy cases.
Questions
Is Claude Haiku 5.5 cheaper than Claude Haiku 4.5?
Per request, yes: it costs about 87% less on six common tasks. Claude Haiku 4.5 lists at $1.00 per million input tokens and $5.00 per million output tokens; Claude Haiku 5.5 lists at $0.10 and $0.50 (Claude Haiku 4.5: price checked 12 nov 2026; unchanged since we started tracking it; Claude Haiku 5.5: price checked 12 nov 2026; unchanged since we started tracking it). Per result that works, it depends on how often each model works first time on your inputs: Claude Haiku 5.5 stays cheaper per result that works as long as it works first time at least 13% as often as Claude Haiku 4.5.
How do I work out the break-even for my own app?
Divide what a request costs on Claude Haiku 5.5 by what it costs on Claude Haiku 4.5, using your real token counts. Claude Haiku 5.5 is cheaper per result that works as long as its share of results that work first time, divided by Claude Haiku 4.5's, is above that number. The model switch test gives you the token counts and cost of each request on both models.
Will Claude Haiku 5.5 work as well as Claude Haiku 4.5 for my app?
Nobody can tell you that from a price table or a general benchmark, because it depends on your prompts and inputs. Run 10 to 20 of your real requests through both models with the free model switch test, using your own provider key, and compare the answers side by side before you switch any users.