NEWEST HAIKU1M CONTEXT90% BELOW HAIKU 4.55x ABOVE 100K PROMPTTEXT + VISIONBATCH -50%
Claude Haiku 5.5 API Pricing
Claude Haiku 5.5 is Anthropic's current Haiku, released October 7, 2026 for classification, routing, extraction and subagent work. Up to a 100,000-token prompt it costs $0.10/M input, $0.50/M output and $0.01/M cache-hit input - 90% below Claude Haiku 4.5. It is the first Claude priced by prompt length: above 100,000 tokens the whole request moves to a card five times higher, $0.50 / $2.50. The window is 1M tokens, thinking is on by default, and the newer tokenizer counts the same text as about 30% more tokens than Haiku 4.5. Checked by hand against platform.claude.com.
Input - per 1M tokens
$0.10/M
Up to 100K prompt -90% vs Haiku 4.5
Output - per 1M tokens
$0.50/M
Includes thinking tokens -90% vs Haiku 4.5
Cached input - 90% off
$0.01/M
Cache write $0.125 / 5 min -90%
Effective - agentic blend
$0.0682/M
92/8 split - 82% cache
§ 01 / TERMINAL
Run the numbers.
Live calculator pre-loaded with Claude Haiku 5.5's up-to-100K rates. Keep an eye on prompt length: a request whose prompt passes 100,000 tokens - counting cache reads and writes - pays the higher card on every token, so for long prompts multiply this calculator's figures by five. Adaptive thinking is on by default at medium effort; put expected reasoning tokens in the reasoning field, since they bill at the output rate.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Real-world presets.
CLASSIFICATION
Ticket routing label
$0.00007/item
SUPPORT AGENT
Live chat first reply
$0.00065/ticket
EXTRACTION
Invoice field extraction
$0.00136/doc
SUBAGENT
Context compaction
$0.00476/task
Estimate · anthropic-bpe-estimate · ≈2.6 chars/token Auto-counts as you type
This is a chars-per-token approximation, not a real tokenizer. Actual tokens vary by language, code density, and tool-call overhead — counts are typically ±10–20% off for English prose, more for code or non-Latin scripts. For exact billing, use the vendor's official tokenizer.
Characters 725
Words 119
Tokens (estimated) 279 tokens
Cost as input · uncached $0.00003 USD
Cost as output · uncached $0.00014 USD
Cost as cached input $0.000003 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Claude Haiku 5.5 Current | $0.10 cache $0.01 | $0.50 | $0.0682 agentic 92/8 | 1M | High-volume classification and subagents |
| Claude Haiku 4.5 | $1.00 cache $0.10 | $5.00 | $0.682 +10× | 200K | The Haiku it replaces, 10x the short-prompt rate |
| Claude Sonnet 5.5 | $2.00 cache $0.10 | $10.00 | $1.29 +19× | 1M | Complex agentic coding, 20x the input rate |
| Claude Opus 5.5 | $4.00 cache $0.20 | $20.00 | $2.58 +38× | 1M | Anthropic's recommended default |
| GPT-6 Luna | $0.10 cache $0.01 | $0.50 | $0.0682 same | 1.05M | OpenAI model at the same short-prompt card |
| Mistral Small 4 | $0.15 cache $0.015 | $0.60 | $0.0842 +23% | 256K | Mistral small multimodal model |
Frequently asked.
Practical Claude Haiku 5.5 pricing questions - the two-card rate sheet, what it means against Haiku 4.5, and where the bill can surprise you.
Q · 01 What is Claude Haiku 5.5 priced at? +
For a prompt up to 100,000 tokens Anthropic lists
claude-haiku-5-5 at $0.10/M input, $0.01/M cache-hit input and $0.50/M output, with cache writes at $0.125/M for 5 minutes and $0.20/M for 1 hour. For a prompt over 100,000 tokens the card is $0.50 input, $0.05 cache hits, $2.50 output, and $0.625 and $1 for the two cache writes.Q · 02 Why does Haiku 5.5 have two sets of prices? +
Anthropic prices it by prompt length, which no other current Claude model does: Claude 4.6 and later models carry the full 1M window at one price, and Haiku 5.5 is the stated exception. The prompt length counts every input token, including cache reads and cache writes. A request over 100,000 tokens pays the higher card on all of its tokens - output included - even when most of the prompt is a cache hit, and each request is priced on its own. Anthropic says about 90% of requests to Haiku 4.5 were under 100,000 tokens, so most traffic lands on the lower card.
Q · 03 How much cheaper is it than Claude Haiku 4.5? +
Per token, 90% cheaper up to 100,000 tokens and 50% cheaper above: Haiku 4.5 lists at
$1/M input and $5/M output with no length tier. Per task the saving is smaller, because Haiku 5.5 uses the newer tokenizer and the same text counts as about 30% more tokens; Anthropic puts the average running-cost cut at about 75%. Under this site's blend the figures are $0.0682/M here against $0.682/M for Haiku 4.5 - before the tokenizer difference.Q · 04 How does it compare with GPT-6 Luna? +
On short prompts the cards are identical: GPT-6 Luna lists at
$0.10/M input, $0.01/M cached and $0.50/M output. The difference is length: Haiku 5.5 moves to $0.50 / $2.50 above 100,000 tokens. In Anthropic's launch table Haiku 5.5 scores above GPT-6 Luna on GDPval-AA (1620 vs 1437 Elo), OSWorld 2.1 offline (72.4% vs 48.9%), Terminal-Bench 4.0 (39.2% vs 16.4%) and FrontierCode (46.4% vs 42.4%) - figures Anthropic ran and reported, not independent results.Q · 05 How is cached input billed? +
Cache hits cost
$0.01/M, the standard 10% of input; the 5% cache-read rate Anthropic introduced applies to Opus 5.5 and Sonnet 5.5, not Haiku. Writes cost $0.125/M for a 5-minute cache and $0.20/M for a 1-hour cache. Cached tokens still count toward the 100,000-token threshold, so caching a long prefix does not keep a request on the lower card.Q · 06 Is there a Batch API discount? +
Yes, 50% on input and output:
$0.05/M and $0.25/M up to 100,000 tokens, and five times those rates above. On the Message Batches API Haiku 5.5 can return up to 300k output tokens with the output-300k-2026-03-24 beta header, against 128k on the synchronous API.Q · 07 Can I turn thinking off? +
Yes, unlike Sonnet 5.5. Adaptive thinking is on by default, so a response can begin with thinking blocks even when the request does not ask for them, and reasoning tokens bill as output.
thinking: {"type": "disabled"} is still accepted at high effort or below, but Anthropic recommends steering cost with the effort parameter instead; the default is medium. Manual budget_tokens returns a 400 error, and so does any non-default temperature, top_p or top_k.Q · 08 What breaks when migrating from Haiku 4.5? +
Anthropic lists five breaking changes: manual extended thinking (
budget_tokens) returns an error; non-default sampling parameters return an error; assistant-message prefill returns an error, so messages must end with a user turn; computer use needs computer_toolset_20260801 on the Claude API and Google Cloud; and editing earlier turns invalidates thinking blocks. Responses can also start with thinking blocks, so read content blocks by type, not position, and recount tokens on your own prompts before budgeting.Q · 09 Does US-only inference change the price? +
Yes. Setting
inference_geo: "us" applies a 1.1x multiplier to every token category on Claude 4.6 and later models, Haiku 5.5 included. The prices on this page use the default global routing. Bedrock and Google Cloud set their own regional prices.Q · 10 What model ID should developers use? +
claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock. Anthropic commits to retirement not sooner than October 7, 2027.