Claude Haiku API pricing: Haiku 5.5 and 4.5
Haiku 5.5 costs $0.10 per 1M input and $0.50 per 1M output tokens for prompts up to 100K tokens (above that: $0.50/$2.50). Haiku 4.5 is $1/$5. A 100,000-item classification run at 500 in / 10 out costs $5.50 on Haiku 5.5.
Nearest-priced models, this workload
Showing the 9 models closest in price; change any input to rank every model.
| Model | Per call | Per 1K calls |
|---|---|---|
| Ministral 3 3B | $0.000051 | $0.0510 |
| Gemini 2.5 Flash-Lite | $0.000054 | $0.0540 |
| GPT-6 Luna | $0.000055 | $0.0550 |
| Claude Haiku 5.5 | $0.000055 | $0.0550 |
| Ministral 3 8B | $0.0000765 | $0.0765 |
| GPT-4o mini (deprecated) | $0.000081 | $0.0810 |
| Mistral Small 4 | $0.000081 | $0.0810 |
| Ministral 3 14B | $0.000102 | $0.1020 |
| GPT-5.6 Luna | $0.000112 | $0.1120 |
o4-mini is shut down; its last published list rate is kept for comparing old invoices, not as a model you can call today — the model page names its replacement.
What Claude Haiku costs on the Anthropic API
Haiku 5.5 is the cheapest Claude model: $0.1 per 1M input tokens and $0.5 per 1M output tokens for prompts up to 100K tokens. Above 100K input tokens the official rate changes to $0.50 input and $2.50 output per 1M. Haiku 4.5, the previous generation, is $1/$5.
| Model | Input /1M | Cached read /1M | Output /1M | Checked |
|---|---|---|---|---|
| Claude Haiku 5.5 | $0.1 | $0.01 | $0.5 | 2026-11-15 |
| Claude Haiku 4.5 | $1 | $0.1 | $5 | 2026-11-15 |
Rates copied from the Anthropic pricing page, checked on the dates shown. The 100K-token boundary is a rate rule from the official page, not a site estimate.
Haiku 5.5 vs Haiku 4.5
Haiku 5.5 is ten times cheaper than Haiku 4.5 on both input ($1 to $0.1) and output ($5 to $0.5), and its cached read is $0.01 against $0.1. For high-volume classification and extraction that difference decides whether a project is viable.
What a Claude Haiku workload costs
Haiku is billed per token like every other model, so the numbers are small until the volume is not. Prices below are per 1,000 calls and per month at the stated volume.
| Workload | Haiku 5.5 per 1K calls | Haiku 5.5 monthly | Haiku 4.5 per 1K calls | Haiku 4.5 monthly |
|---|---|---|---|---|
| Classify or tag one item (500 in / 10 out) | $0.0550 | $5.50 | $0.5500 | $55.00 |
| Generate a product title (200 in / 30 out) | $0.0350 | $1.75 | $0.3500 | $17.50 |
| Route a request before a bigger model (1K in / 50 out) | $0.1250 | $125 | $1.25 | $1,250 |
Routing is the one workload where the comparison is not Haiku against Haiku: at 1M calls a month, 1K in / 50 out costs $125 on Haiku 5.5 against $2,500 on Sonnet 5.5 for the same tokens, before you count the Sonnet calls the routing avoids.
When to use Haiku instead of Sonnet
Use Haiku where the output is short and the task is narrow: tags, categories, yes/no gates, title and meta generation, extraction from a fixed schema, and routing in front of a bigger model. Escalate to Sonnet 5.5 when the answer needs multi-step reasoning or a long, unstructured response; the price difference only pays off if the cheap model actually finishes the job.
Haiku 5.5 at volume, including the 100K context step
Haiku 5.5 lists $0.1 per 1M input and $0.5 per 1M output tokens, with cache reads at $0.01. The cheap rate card is also where long context stops being a small line item: prompts over 100,000 tokens switch the whole request to the long-context rate in our specs file ($0.5 input, $2.5 output).
A 10,000-token prompt with a 1,000-token answer
One call costs $0.0015 and 1,000 calls cost $1.50. Output is 33% of that. Claude Haiku 4.5, the version Haiku 5.5 replaced, lists $1/$5: $0.0150 for the same call, 10x the price of the current model on input and the same multiple on output.
A retrieval answer: 5,000 context tokens in, 2,000 out
This costs $0.0015 per answer and $1.50 per 1,000. The prompt stays under the long-context threshold, so the standard rate applies. Gemini 3.1 Flash-Lite lists $0.25/$1.5, which is more on both sides ($0.0040 for the 10,000/1,000 call), and it is the closest cheap-tier comparison on the site.
A 40-turn agent session
A 60,000-token prefix written to cache once, read cached on 39 later turns, 3,000 fresh input and 2,000 output tokens per turn: $0.0826 per session, $82.60 per 1,000 sessions, output 1210654% of it. That is well under the long-context threshold at every turn. The same session on Haiku 4.5 costs $0.8260; on Claude Sonnet 5.5 it is $0.0000 plus the turns: a Sonnet 5.5 session of the same shape is far higher, and the exact figure is on the Sonnet price page.
What happens at 120K input tokens
Cross the threshold once and the whole request reprices: a 120,000-in / 2,000-out call is $0.0130 at the standard rate and $0.0650 at the long-context rate on this model, 5x the input price and 5x the output price. If a workload routinely sends 100K+ prompts, price it at the long-context rate before comparing Haiku with anything else.
Cost per user at cheap-tier volume
At 2,000 calls a month on the 10,000/1,000 mix one user costs $3.00 in model time; at an 80% margin the minimum price is $15.00 per user per month. The same 2,000 calls cost $60.00 on Sonnet 5.5 and $30.00 on Haiku 4.5, which is why the tier choice, not the prompt, sets the price floor on high-volume products.
Haiku price checks since 4 November 2026
The official Anthropic pricing page has been fetched and diffed on 101 completed checks since our monitoring began on 2026-11-04, and none of those checks flagged a change to the list prices above. The row used here was last confirmed against that page on 2026-11-15. Failed fetches are skipped rather than counted, so the number is checks, not days.
For short, high-volume work the batch guide matters more than caching: if an answer nobody is waiting for can run asynchronously, the batch rate halves it. The AI SaaS pricing guide shows how a cheap per-token rate still needs a usage cap, because volume rather than rate sets the monthly bill.
Haiku 5.5 against Haiku 4.5 and the cheap tier
Haiku 5.5 is a step change: Claude Haiku 4.5 lists $1/$5, ten times the Haiku 5.5 rates, and the session above is $0.8260 on 4.5 against $0.0826 on 5.5. Gemini 3.1 Flash-Lite lists $0.25/$1.5, and Claude Sonnet 5.5 lists $2/$10, so the order on cheap work is Haiku 5.5, Gemini 3.1 Flash-Lite, then Sonnet 5.5 at $0.0300 for the 10,000/1,000 call. Head-to-head pages: Claude Haiku 5.5 vs Gemini 3.1 Flash-Lite, GPT-6 Luna vs Claude Haiku 5.5 and GPT-5.4 mini vs Claude Haiku 4.5. For the full per-user arithmetic use the AI SaaS pricing guide and the unit economics calculator; Claude cost calculator prices your own token mix.
Related Claude prices
- Claude Sonnet pricing
- Claude Opus pricing
- Claude Haiku 4.5 pricing (kept as a version page)
- All Claude models on the Anthropic API
- Claude cost calculator (all models)
Questions
How much does Claude Haiku cost per 1M tokens?
Haiku 5.5 is $0.1 input and $0.5 output per 1M tokens for prompts up to 100K tokens; longer prompts move to $0.50/$2.50 per 1M. Haiku 4.5 is $1/$5.
What is the difference between claude haiku pricing and claude haiku api cost?
Pricing is the rate card: $0.10 in and $0.50 out per 1M on Haiku 5.5. Cost is that rate times your tokens, so a classification run of 100,000 items costs what the table above says, not what the headline rate suggests.
Does Haiku 5.5 get more expensive on long prompts?
Yes. For prompts over 100K input tokens Anthropic lists $0.50 per 1M input and $2.50 per 1M output. Below that boundary every token is billed at $0.10/$0.50.
When should I use Haiku instead of Sonnet?
When the task is narrow and the output short: tags, categories, gates, fixed-schema extraction and routing. Sonnet 5.5 is twenty times the Haiku 5.5 input rate, so keep multi-step reasoning and long answers on Sonnet.
Is Haiku good enough for production?
For classification, extraction and routing, yes, and the price makes high volume practical. Check the failure mode before you commit: if the cheap model returns plausible-but-wrong labels, the saved tokens disappear into manual review, and the calculator above is the wrong tool for that decision.
Next step → Estimate cost per user with Claude Haiku
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.