Claude Haiku 5.5 API pricing
Below 100K input tokens, Anthropic lists $0.10 input and $0.50 output per million; this guide shows what the tier and a half-price quote actually mean.
Claude Haiku 5.5 API pricing has two prompt-size bands. For a prompt up to 100K tokens, Anthropic lists $0.10 per million input tokens, $0.50 per million output tokens, and $0.01 per million cache-read tokens. A prompt over 100K uses $0.50 input, $2.50 output, and $0.05 cache read. A 50% calculation on the first band is $0.05 input and $0.25 output, but that arithmetic is not a claim about a particular gateway's live quote.
The useful budgeting question is not only “what is the rate?” It is “which band will this request use, which tokens are cache reads, and does the route actually list this model?” This page owns those calculations. The Claude Haiku 5.5 review covers benchmarks, model behavior, and the breaking changes from Haiku 4.5.
Prices checked October 8, 2026. The provider rates and model limits below come from Anthropic's announcement, model overview, migration guide, and pricing documentation. Half-price figures are explicit arithmetic examples; OmniaKey's live route is priced 20% below this rate card at $0.08 input and $0.4 output per million up to 100K tokens. We did not run a paid benchmark; verify the live model catalog before budgeting a gateway.
Claude Haiku 5.5 pricing at a glance
All amounts below are USD per one million tokens. “Up to 100K” and “over 100K” are rate bands selected by prompt size.
| Item | Prompt up to 100K | Prompt over 100K |
|---|---|---|
| Uncached input | $0.10 / MTok | $0.50 / MTok |
| Output | $0.50 / MTok | $2.50 / MTok |
| Cache read | $0.01 / MTok | $0.05 / MTok |
| Five-minute cache write | $0.125 / MTok | $0.625 / MTok |
| One-hour cache write | $0.20 / MTok | $1.00 / MTok |
| Batch input and output | 50% discount | 50% discount |
| Context window | 1M tokens | 1M tokens |
| Maximum synchronous output | 128K tokens | 128K tokens |
The 1M context window describes the maximum request capacity, not one million free tokens. Anthropic's model overview also lists adaptive thinking, a default API effort of medium, and up to 300K output tokens on the Message Batches API beta path with the documented header. Those limits describe separate contracts; do not use the Batch ceiling as the normal synchronous limit.
What the 100K tier means
The threshold is about the prompt sent to the model. The prompt includes the conversation or system text your client sends, plus any tool definitions and other input that the provider counts. A short visible answer does not make a large prompt a small request.
Read the price card as two bands:
- At or below 100,000 input tokens: use $0.10 input and $0.50 output per million.
- Above 100,000 input tokens: use $0.50 input and $2.50 output per million.
Do not silently treat the first 100K as cheap and only the remainder as expensive unless the provider's billing contract says the tier is marginal. The documentation presents prompt-size rates, so a budget model should classify the request first and then apply the matching band.
| Prompt shape | Band to model | Direct rate calculation |
|---|---|---|
| Exactly 100K input + 20K output | Up to 100K | 0.1 × $0.10 + 0.02 × $0.50 = $0.02 |
| 100,001 input + 20K output | Over 100K | 0.100001 × $0.50 + 0.02 × $2.50 ≈ $0.10 |
| 200K input + 20K output | Over 100K | 0.2 × $0.50 + 0.02 × $2.50 = $0.15 |
A tool-heavy agent can cross the boundary through repository context, tool schemas, or conversation history even when its final answer is only a few paragraphs. Record provider-reported usage rather than estimating from characters or file size.
What “half price” calculates to
A half-price table is useful when a reseller, internal budget, or negotiated route says it charges 50% of Anthropic's standard rate. It must be labeled as a calculation because provider list prices, gateway quotes, cache policies, and eligibility are separate contracts.
| Usage | Anthropic list, up to 100K | 50% arithmetic, up to 100K | Anthropic list, over 100K | 50% arithmetic, over 100K |
|---|---|---|---|---|
| Input | $0.10 | $0.05 | $0.50 | $0.25 |
| Output | $0.50 | $0.25 | $2.50 | $1.25 |
| Cache read | $0.01 | $0.005 | $0.05 | $0.025 |
The formula is simple:
half-price rate = Anthropic list rate × 0.5
That formula does not establish that an account can buy the model at half price. A live quote may use a different ratio, add a platform fee, omit cache pricing, restrict regions, or not expose Haiku 5.5 at all. Check the exact model ID and billing record on the route you will call.
Three reproducible cost examples
These examples isolate token charges. They exclude cache-write charges, taxes, tools, retries, regional multipliers, and any gateway fee.
| Workload | Direct Anthropic calculation | Direct total | 50% arithmetic total |
|---|---|---|---|
| 100K uncached input + 20K output | 0.1 × $0.10 + 0.02 × $0.50 | $0.02 | $0.01 |
| 200K uncached input + 20K output | 0.2 × $0.50 + 0.02 × $2.50 | $0.15 | $0.075 |
| 100K cache read + 20K output | 0.1 × $0.01 + 0.02 × $0.50 | $0.011 | $0.0055 |
The third row assumes the provider reports all 100K input tokens as cache reads. Repeating similar text does not prove a cache hit. Cache creation is a separate charge, and a cache miss uses the uncached input rate.
The first row is also a reminder about unit conversion: 100K tokens is 0.1 million and 20K is 0.02 million. Keep the units visible in a spreadsheet or budget script so a 1,000-fold error is easy to spot.
Cache, Batch, and thinking costs
Prompt caching changes the input line item; it does not make output free. Haiku 5.5 lists a five-minute cache write at 1.25 times the base input rate and a one-hour write at twice the base input rate. The cache-read rates in the table are the provider's separate hit prices.
Batch processing is a different service path. Anthropic documents a 50% discount for Batch input and output, while the model overview separately documents the 300K beta output ceiling. Confirm that your client sends the correct beta header and that your workload can tolerate asynchronous completion.
Adaptive thinking is on by default for Haiku 5.5. Use the usage fields from the response and provider invoice when you estimate an agent task; visible answer length alone does not describe all work performed by a reasoning model. For migration details, including the removal of old sampling parameters and assistant prefills, use the full review.
A minimal request shape
The model ID is claude-haiku-5-5. The following body shows the adaptive-thinking contract documented for the Claude API. Send it only through a provider or gateway route that lists this exact model ID.
{
"model": "claude-haiku-5-5",
"max_tokens": 4096,
"thinking": {"type": "adaptive"},
"output_config": {"effort": "medium"},
"messages": [
{"role": "user", "content": "Extract the three risks and return JSON."}
]
}
For a gateway, the endpoint and authentication header come from that gateway's API quick start. Keep the direct Anthropic price card separate from the gateway's current quote, and log the requested model plus returned usage so you can reconcile a bill.
How to check a gateway quote
- Search the live catalog for
claude-haiku-5-5, not only for the “Haiku 5.5” display name. - Confirm input, output, cache-read, long-context, Batch, and regional rules separately.
- Run one bounded request with a scoped key and record the returned usage fields.
- Compare cost per accepted task, including failed retries and human correction, rather than comparing one token rate.
OmniaKey's public catalog now lists Haiku 5.5 as claude-haiku-5-5, priced 20% below Anthropic's rate card: $0.08 input and $0.4 output per million up to 100K tokens, and $0.4 input and $2 output above it. The model catalog is the source for current OmniaKey availability and pricing; the $0.05/$0.25 arithmetic above remains a hypothetical, not an OmniaKey promise.
For benchmarks, latency context, and Haiku 4.5 migration hazards, read the Haiku 5.5 review. For a broader comparison of Claude and OpenAI billing, see Claude API versus OpenAI API pricing.
Frequently asked questions
How much does Claude Haiku 5.5 cost?
For prompts up to 100K tokens, Anthropic lists $0.10 per million input tokens, $0.50 per million output tokens, and $0.01 per million cache-read tokens. Above 100K, the corresponding rates are $0.50, $2.50, and $0.05.
Is $0.05 input and $0.25 output the official price?
No. Those numbers are exactly half of the up-to-100K Anthropic rates. They are valid arithmetic for a route that explicitly charges 50% of list, not proof that Anthropic or a particular gateway offers that quote.
Does the 100K tier make the first 100K tokens cheap?
The documentation presents prompt-size bands. Treat a prompt over 100K as an over-100K request unless your billing contract explicitly defines marginal tiers. Confirm the rule before implementing a cost calculator.
Does a 1M context window include free usage?
No. The 1M figure is capacity. Input, cache activity, and output remain metered under the applicable rate band.
Why can a 50% token rate produce less than 50% task savings?
Haiku 5.5 uses a newer tokenizer that produces approximately 30% more tokens for the same text than Haiku 4.5, and retries, tools, thinking, cache writes, and gateway fees can change the total. Anthropic reports average task-cost results separately; do not turn one rate-card calculation into a universal saving claim.
Is Claude Haiku 5.5 available through OmniaKey?
Yes. OmniaKey lists claude-haiku-5-5 at 20% below Anthropic's official rate card, from $0.08 input / $0.4 output per million up to 100K tokens. Check the current model catalog for live pricing.