AI Tools
Claude Haiku pricing: what does Haiku 5.5 cost per million tokens?
Claude Haiku pricing now starts at $0.10 per million input tokens and $0.50 per million output tokens, after Haiku 5.5 launched on October 7, 2026. Cross 100,000 prompt tokens and both rates jump fivefold, to $0.50 and $2.50, which is the detail most Haiku cost estimates leave out.
What matters
- Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, per Anthropic's pricing docs.
- The same model costs $0.50 and $2.50 per million above 100,000 prompt tokens. Anthropic says 90% of requests to Haiku 4.5 stayed under that line.
- The Batch API halves both rates to $0.05 and $0.25, and cached reads cost $0.01 per million under the threshold.
- Anthropic prices Haiku 5.5 90% below Haiku 4.5 under 100k tokens, but its own footnote puts the blended saving at roughly 75% because the new tokenizer spends slightly more tokens per task.
- Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per million tokens on the same day, which Anthropic says cuts its cost on most agentic tasks by around 20%.
Claude Haiku 5.5 pricing per million tokens
Anthropic released Claude Haiku 5.5 on October 7, 2026 under the model ID claude-haiku-5-5, and it is the first Haiku-class model with two published rate tiers. Per Anthropic's pricing documentation, prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Prompts over that boundary cost $0.50 and $2.50 per million.
Cache pricing follows the same split. A cached read costs $0.01 per million tokens under the threshold and $0.05 above it. A five-minute cache write costs $0.125 per million under and $0.625 over, and a one-hour write costs $0.20 or $1.00. Those write rates are 1.25x and 2x the base input rate, so caching only pays if you read the same context back repeatedly.
The rest of Anthropic's current line, from the same page: Haiku 4.5 at $1.00 input and $5.00 output per million, Sonnet 5.5 at $2.00 and $10.00, Opus 5.5 at $4.00 and $20.00, and the Fable 5.1 flagship at $10.00 and $50.00. Haiku 5.5 is the cheapest of the group on both input and output.
What the 100,000-token cliff actually costs
The threshold applies to the prompt, so it is a property of each request rather than of a whole session. A 150,000-token prompt, the size you would feed a codebase-wide review or a full contract set, costs $0.015 of input at the low rate and $0.075 once it crosses. A 10,000-token answer generated from that prompt costs $0.005 at the low rate and $0.025 at the high one. The step is exactly five times, in both directions.
Two things soften it. First, the high tier is still cheap in absolute terms: $0.50 per million input tokens is four times less than Sonnet 5.5's $2.00, and $2.50 output is four times less than Sonnet's $10.00. Crossing the line narrows Haiku 5.5's advantage over Sonnet from 20x to 4x, it does not make the small model the expensive choice.
Second, Anthropic states that prompts under 100,000 tokens covered around 90% of requests to the previous Haiku model, which is why the tier structure is set where it is. Chat, classification, extraction, routing and compaction workloads built on documents or support tickets sit comfortably below the boundary. Pipelines that stuff a whole repository into every call do not.
The 90% price cut is real, the blended saving is about 75%
Anthropic's launch page puts the comparison to Haiku 4.5 at 90% lower for requests up to 100,000 tokens and 50% lower above that, with an average of around 75% less to run. The footnote explains why the two numbers disagree: Haiku 5.5 ships with an updated tokenizer, similar to the ones in Sonnet 5.5 and Opus 5.5, and it uses slightly more tokens to complete the same task.
For anyone modelling spend, that means the per-token price is the headline and the token count is the correction. If your workload currently burns 200 million tokens a month on Haiku 4.5, the same work on Haiku 5.5 lands near 215 to 225 million tokens before any price maths, and the bill still collapses from roughly $400 to under $50. The tokenizer drift is not a reason to stay on the old model, only a reason not to quote the 90% figure as your final answer.
Batch and caching: the two levers that decide the bill
Batch API. Asynchronous work gets 50% off input and output, which puts Haiku 5.5 at $0.05 and $0.25 per million under the threshold, and $0.25 and $1.25 above it. If your job can wait, overnight batch jobs on Haiku 5.5 cost less per million tokens than most open-weight models hosted elsewhere.
Prompt caching. A cache read costs $0.01 per million tokens on Haiku 5.5 under the threshold, ten times cheaper than Haiku 4.5's $0.10. A 10-million-token shared context read 20 times a month, so 200 million read tokens, costs $2 on Haiku 5.5 and $20 on Sonnet 5.5 at its new $0.10 rate. The trade-off is at write time: you pay 1.25x the input rate for a five-minute cache and 2x for an hour, so a context you reuse twice is a loss and one you reuse all day is close to free.
One honest catch on caching: the multipliers are charged against the base input rate, so above 100,000 tokens a five-minute cache write costs $0.625 per million rather than $0.125. Long-context caching on Haiku 5.5 is cheap to read and expensive to build.
Sonnet 5.5 got cheaper on the same day
The Haiku launch carried a second price change: Anthropic halved Claude Sonnet 5.5's cache reads from $0.20 to $0.10 per million tokens, and says that reduces Sonnet 5.5's cost on most agentic tasks by around 20%. Cache reads are the bulk of token consumption in agent loops that re-send the same file or conversation state, so the cut lands hardest exactly where the mid-tier model is used.
Put together, the two changes move both ends of the ladder at once. A build-versus-buy decision that chose Haiku for cost only is now less clear-cut: a Sonnet 5.5 agent whose spend is mostly cache reads moved closer to the small model without any drop in capability. The clean way to compare is to price your own token mix. A pipeline running 240 million input and 40 million output tokens a month costs $44 on Haiku 5.5, $440 on Haiku 4.5 and $880 on Sonnet 5.5, all at standard tier and all before caching or batch discounts.
Who should pay for Haiku 5.5, and who should stay on Sonnet
Anthropic's own guidance is more conservative than the launch headline. The company states that Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding, and its published benchmark table backs that up: Haiku 5.5 scores 39.2% on Terminal-Bench 4.0 against Sonnet 5.5's 70.6%, and 1620 against 1840 on the GDPval-AA v2.1 knowledge work set. Against Haiku 4.5 the same table shows 39.2% versus 0.0% on Terminal-Bench 4.0 and 72.4% versus 15.7% on the OSWorld 2.1 computer-use subset.
- Buy Haiku 5.5 for classification, extraction, routing, summarisation, compaction and subagent steps inside a larger pipeline, where the work is narrow and the volume is high. The 20x price gap is real and the quality drop rarely shows on short, well-scoped tasks.
- Stay on Sonnet 5.5 for agentic coding, long tool-use chains, or anything where one bad step wastes the run. Paying eight times more per output token is cheaper than retrying a failed session with the small model.
- Re-test Haiku 4.5 before renewing on price alone. At $1.00 and $5.00 per million it now costs ten times Haiku 5.5 under the threshold, so your eval suite is the only question left.
Haiku 5.5 also carries a 1-million-token context window, a 128,000-token maximum output, a June 2026 knowledge cutoff, and a published retirement date no sooner than October 7, 2027. It is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Max and Team subscribers now get monthly API credits
Alongside the rate card, Anthropic says it is rolling out a monthly Claude Platform credit for consumer and team subscribers. Max 5x users get $100 per month, Max 20x users get $200, and Team subscribers receive up to $500 pooled across their seats. The credits can be spent on any Claude model through the API, which means a subscriber running a Haiku 5.5 side project can cover 200 million input tokens a month out of the allowance at the low tier.
For a buyer weighing Claude against a competing API on price, that changes the effective cost of the first experiment. The API list prices above are what you pay once the credits run out, not what a Max or Team account pays for its first project of the month.
At a glance
| Rate per 1M tokens | Haiku 5.5 (up to 100k) | Haiku 5.5 (over 100k) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 | $2.00 |
| Output | $0.50 | $2.50 | $5.00 | $10.00 |
| Cache read | $0.01 | $0.05 | $0.10 | $0.10 |
| 5-minute cache write | $0.125 | $0.625 | $1.25 | $2.50 |
| Batch input / output | $0.05 / $0.25 | $0.25 / $1.25 | $0.50 / $2.50 | $1.00 / $5.00 |
| Best fit | High-volume narrow tasks and subagents | Long-context pipelines | Previous Haiku generation | Complex agentic coding |
FAQ
How much does Claude Haiku 5.5 cost per million tokens?
Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and $0.50 input with $2.50 output above that. Cache reads are $0.01 and $0.05 per million, and the Batch API halves input and output to $0.05 and $0.25 under the threshold, per Anthropic's pricing docs.
Is Claude Haiku 5.5 cheaper than Claude Haiku 4.5?
Yes. Anthropic prices Haiku 5.5 90% below Haiku 4.5 for prompts up to 100,000 tokens and 50% below for prompts over 100,000 tokens. The company's own footnote puts the blended saving at around 75% because Haiku 5.5 uses an updated tokenizer that spends slightly more tokens on the same task.
Do prompts over 100,000 tokens make Haiku 5.5 more expensive than Sonnet 5.5?
No. Haiku 5.5's over-100k rates of $0.50 input and $2.50 output per million tokens are still four times cheaper than Sonnet 5.5 at $2.00 and $10.00. The high tier narrows the gap from 20x to 4x rather than closing it, and batch or cached work widens it again.
Related reading
Claude tool profile, Claude API pricing across the model line, Cheapest LLM API