Claude Haiku 5.5 Pricing

Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens. A prompt of over 100,000 tokens pays higher prices: $0.50 input and $2.50 output. Cache reads start at $0.01, the Batch API halves every rate, and Anthropic says the model costs “around 75% less to run” than Haiku 4.5 on average.

Try Apidog today

This guide covers every line item, the math behind that 75%, three worked examples (one crossing 100K), cost per effort level, and how Haiku 5.5 compares with Haiku 4.5, GPT-6 Luna and Sonnet 5.5. New to the model? Start with what Claude Haiku 5.5 is. To check the numbers on your own traffic, send requests from Apidog and read each response’s usage block.

Claude Haiku 5.5 API pricing table

Prices per million tokens (MTok), from Anthropic’s pricing docs:

Line item	Prompts up to 100K tokens	Prompts over 100K tokens
Input	$0.10	$0.50
Output	$0.50	$2.50
5-minute cache write	$0.125	$0.625
1-hour cache write	$0.20	$1.00
Cache read (hits and refreshes)	$0.01	$0.05
Batch input	$0.05	$0.25
Batch output	$0.25	$1.25

The docs put it plainly: “a prompt of over 100,000 tokens pays higher prices.” Other Claude models from 4.6 onward bill the full 1M window at one rate; Haiku 5.5 is the exception.

Also account for the following before estimating production cost:

• US-only inference. US-only routing via inference_geo costs 1.1x on every token category on the Claude API and Claude Platform on AWS: $0.11 input and $0.55 output under 100K.
• A new tokenizer. The same text produces about 30% more tokens than on Haiku 4.5.
• Smaller tool overhead. The tool-use system prompt is 286 tokens (auto/none), down from 496 on Haiku 4.5.
• No Priority Tier and no fast mode.

Where “around 75% less” comes from

Haiku 5.5 is priced 90% lower than Haiku 4.5 up to 100,000 tokens and 50% lower above. Anthropic says 90% of Haiku 4.5 requests fell under 100K, and its estimate also accounts for the new tokenizer.

A rough reconstruction:

1. Weight the reduction by request count:

   0.9 × 90% + 0.1 × 50% = 86% lower

1. Account for approximately 30% more tokens:

   Remaining cost: 14%
   Tokenizer-adjusted cost: 14% × 1.3 = 18.2%
   Savings: about 82%

Anthropic’s figure is lower again. It does not publish the weighting, and long prompts likely represent more spend than request count alone suggests. Treat 75% as an average; your own usage data is the real answer.

Cost per attempt at each effort level

Haiku 5.5 is the first Haiku with effort levels: low through max, with medium as the API default. Effort can affect your bill more than the base rate card.

These figures come from the launch post’s per-effort charts. Anthropic ran OSWorld and Terminal-Bench; Artificial Analysis ran GDPval-AA.

Effort	OSWorld 2.1 (score, $/attempt)	GDPval-AA v2.1 (Elo, $/task)	Terminal-Bench 4.0 (score, $/attempt)
low	42.0%, $0.0695	1125, $0.0117	12.7%, $0.424
medium	53.3%, $0.1257	1277, $0.0304	20.3%, $0.6798
high	61.3%, $0.1827	1420, $0.0894	24.8%, $1.0372
xhigh	67.6%, $0.2792	1513, $0.2673	31.5%, $1.7545
max	72.4%, $0.6111	1620, $0.8659	39.2%, $2.6433

Use the table to select effort based on the task, not just benchmark scores:

• Max costs 4.9x medium on OSWorld ($0.6111 ÷ $0.1257) and 28.5x medium on GDPval-AA ($0.86593 ÷ $0.03042).
• Launch-table scores are max effort. At the default medium, GDPval-AA is 1277, not 1620.
• For agentic coding, check Sonnet. On Terminal-Bench, Sonnet 5.5 at low with $0.10 cache reads scored 20% for $0.6205. Haiku 5.5 at medium scored 20.3% for $0.6798. Anthropic says Sonnet and Opus “remain better choices for complex agentic coding tasks.” See the full benchmarks breakdown.

How Haiku 5.5 compares on price

Model	Input / output per MTok	Cache read	5m cache write	Long-prompt rule
Claude Haiku 5.5	$0.10 / $0.50 up to 100K	$0.01	$0.125	Over 100K: $0.50 / $2.50
GPT-6 Luna	$0.10 / $0.50	$0.01	$0.125	Over 272K input: 2x input and cache, 1.5x output ($0.20 / $0.75)
Claude Haiku 4.5	$1 / $5	$0.10	$1.25	One rate (200K context)
Claude Sonnet 5.5	$2 / $10	$0.10	$2.50	One rate across 1M

GPT-6 Luna matches Haiku 5.5 up to 100K, but the long-context threshold differs. Per OpenAI’s model page, the Luna surcharge starts above 272K and applies “for the full request.”

For example, a request with 150,000 input tokens and 2,000 output tokens stays on Luna’s base rate:

(150,000 × $0.10 + 2,000 × $0.50) ÷ 1,000,000 = $0.016

The same request costs $0.08 on Haiku 5.5 because it crosses Haiku’s 100K threshold.

Tokenizers differ too: one Hacker News commenter estimated 100K Claude tokens at about 60–65K GPT tokens. On Anthropic’s OSWorld chart, Haiku 5.5 at medium (53.3%, $0.1257) beat Luna at max (48.9%, $0.205). See Claude Haiku 5.5 vs GPT-6 Luna for more detail.

Sonnet 5.5 cut cache reads from $0.20 to $0.10 the same day, making it about 20% cheaper on most agentic work per Anthropic. The Sonnet 5.5 pricing guide predates the cut.

Haiku 4.5 is not deprecated. Review Haiku 5.5 vs Haiku 4.5 for breaking changes before migrating.

API credits and consumer plans

Max and Team API credits. Max 5x gets $100 a month in Claude Platform credits, Max 20x gets $200, and Team gets $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. Free, Pro, and Enterprise are not eligible.

Per the API credits docs, credits cover the Claude API, not Claude Code or cloud providers, and they do not roll over.

At $0.10 per MTok, $100 buys 1 billion input tokens, or about 220,000 Example 1 requests:

$100 ÷ $0.000455 ≈ 220,000

Consumer plans. According to claude.com/pricing, Free ($0), Pro, Max, Team, and Enterprise users can select Haiku 5.5 on Claude.ai on web, iOS, and Android. Pro is $17/month billed annually or $20 monthly; Max starts at $100/month.

Free has no Claude Code, and a chat plan cannot drive a script or CI job. See how to use Claude Haiku 5.5 for free for the remaining options.

Clouds. Bedrock, Google Cloud, and Foundry bill Haiku 5.5 through their own price pages.

Track cost per request in Apidog

Every Messages response returns a usage object. You can use it to calculate per-request cost in Apidog.

1. Store your key as an ANTHROPIC_API_KEY environment variable.
2. Save one request per effort level—low, medium, and high—while keeping the prompt constant.

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-haiku-5-5",
    "max_tokens": 4000,
    "thinking": {"type": "adaptive"},
    "output_config": {"effort": "medium"},
    "messages": [{"role": "user", "content": "Classify this support ticket: ..."}]
  }'

1. Add a post-processor that selects the correct pricing tier and saves the request cost. This assumes no caching, so input_tokens represents the prompt length.

const body = pm.response.json();
const u = body.usage;

const long = u.input_tokens > 100000;
const rateIn = long ? 0.50 : 0.10;
const rateOut = long ? 2.50 : 0.50;

const cost = (u.input_tokens * rateIn + u.output_tokens * rateOut) / 1e6;

pm.environment.set("haiku_call_usd", cost.toFixed(6));

pm.test("not a refusal", () => {
  pm.expect(body.stop_reason).to.not.eql("refusal");
});

1. Run all three saved requests and compare output tokens and cost side by side.

The refusal check matters because Haiku 5.5’s safety classifiers have no server-side fallback. See how to use the Claude Haiku 5.5 API for the full walkthrough.

FAQ

How much does Claude Haiku 5.5 cost per million tokens?

$0.10 input and $0.50 output for prompts up to 100K tokens; $0.50 input and $2.50 output above 100K.

Is Haiku 5.5 cheaper than Haiku 4.5?

Yes. It is 90% cheaper per token up to 100K, 50% cheaper above 100K, and around 75% cheaper on average per Anthropic.

Is Haiku 5.5 the same price as GPT-6 Luna?

Up to 100K, yes. Above that, Haiku 5.5 rises 5x, while Luna stays flat until 272K. See the GPT-6 Luna overview for details.

Do Max plan credits work in Claude Code?

No. They cover the Claude Platform API, not Claude Code. See Claude Haiku 5.5 in Claude Code for plan access there.

Next step: price your own workload

Pull your three most common request shapes, check how many cross 100K, and run each at low and medium. Multiply the returned usage values by the pricing table to estimate your real savings.

Download Apidog to keep those requests and the cost script in one project.