Claude Haiku 5.5 Guide: Pricing, Benchmarks, Rate Limits and When to Use It

Claude Haiku 5.5 is Anthropic’s fastest and cheapest model in the 5.5 lineup, launched on October 7, 2026.

It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, has a 1M token context window, and on Artificial Analysis’s new index it outscores every open-weight model you can download.

Here is everything you need to decide if it fits your app, plus code and rate limits.

What Is Claude Haiku 5.5?

Haiku is Anthropic’s small, fast model family. Haiku 5.5 is the newest member, and Anthropic describes it as its most capable model tuned for high-volume and latency-sensitive work. Think classification, routing, data extraction, chat, and subagents that run many small tasks inside a bigger agent.

It is a big jump from Haiku 4.5. The context window grew from 200K to 1M tokens, max output doubled from 64K to 128K, and it is the first Haiku with adaptive thinking and the effort parameter. Artificial Analysis also noted that its base price is about 10% of the previous Haiku’s price.

Haiku 5.5 is available on the Claude API, Claude in Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud and Claude in Microsoft Foundry. On Bedrock the model id is anthropic.claude-haiku-5-5. Everywhere else it is claude-haiku-5-5.

Haiku 5.5 Pricing Explained

Haiku 5.5 is the only current Claude model with two price tiers based on prompt length. Every other Claude 4.6 or later model charges the same rate across the full 1M context. Here are the official numbers from Anthropic’s pricing page, in US dollars per million tokens.

Three rules decide which tier you pay:

• The 100K line counts all input tokens, including cache reads and cache writes.
• Pricing is per request. If a request goes over 100K, the whole request pays the higher rates, even the cached part.
• US-only inference (inference_geo: "us") adds a 1.1x multiplier, including on the higher tier.

Two hidden costs also matter. First, Haiku 5.5 uses a newer tokenizer that produces about 30% more tokens than Haiku 4.5 for the same text. Second, adaptive thinking is on by default and thinking tokens are billed as output. Lower the effort level if you do not need deep reasoning.

A quick example. Say you run 1 million requests, each with 2,000 input tokens and 300 output tokens. On Haiku 5.5 that is about $200 for input plus $150 for output, so $350.

The same job on Sonnet 5.5 would be about $7,000, and on Opus 5.5 about $14,000. These are list prices before caching, batching or thinking tokens. For the full Claude price list, see my Claude API pricing guide.

Context Window and Specs

Tip: The lower 512-token caching minimum is handy. Short system prompts that could not be cached on Haiku 4.5 can now be cached for $0.01 per million tokens on a read. My prompt caching guide shows how to set it up.

Max output with the 300K Batches option needs the output-300k-2026-03-24 beta header. Anthropic says Haiku 5.5 will not retire sooner than October 7, 2027.

Benchmarks

The most cited numbers come from Artificial Analysis, an independent group that tests models on its own suite.

At launch (October 7, 2026). Artificial Analysis scored Haiku 5.5 at 43 on its Intelligence Index at max effort, up 26 points from the previous Haiku. At high effort it scored 38.

It reported 33% on Terminal-Bench 4.0, up from 0% for Haiku 4.5, and a hallucination rate of 40% on AA-Omniscience, lower than Gemini 3.8 Flash (55%) and GPT-6 Luna (77%).

One weak spot: it scored 35% on AutomationBench-AA, behind rivals at 53 to 60%, partly due to an over-refusal issue Anthropic reported in pre-release testing.

After Index v5 (late October 2026). Artificial Analysis rebuilt its index around private test sets, with about 60% of it now held out. On this new version, Claude Opus 5.5 leads with 58 and Haiku 5.5 scores 36.

Business Insider reported in November 2026 that the best open-weight models, whose weights anyone can download, scored 32. That is how Haiku 5.5, Anthropic’s smallest model, ended up ahead of every downloadable open-weight model on the board.

A fair caveat: the open-weight gap depends on how you count. Kimi K3.1 Pro scored 46, but its weights have not been released, so it is not counted as downloadable. Haiku 5.5 is also verbose.

Artificial Analysis found it used about 162K output tokens per index task at max effort, roughly three times GPT-6 Luna. For live rankings across all vendors, see the best AI model leaderboard for November 2026.

Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5

All three share a 1M context and 128K max output. The differences are price, speed and depth.

Pick Haiku 5.5 when: you run thousands or millions of short requests, latency matters (live chat, autocomplete, voice), the task is well defined (tagging, classifying, pulling fields out of text), or you need a cheap subagent that does one job inside a larger Sonnet or Opus agent.

Pick Sonnet 5.5 when: prompts often go above 100K tokens. At that point Haiku costs $0.50 / $2.50 and Sonnet’s flat $2 / $10 is only four times higher, while Sonnet is a much stronger model. Also pick Sonnet for serious coding and agents that must plan across many steps.

Pick Opus 5.5 when: accuracy on hard problems matters more than cost, such as deep research, tricky code changes or high-stakes analysis.

A common pattern is to mix them: Haiku triages and routes, Sonnet or Opus handles the hard cases. Your cost report will show if the split is working. To compare with other providers’ small models, see the AI API pricing comparison.

How to Use Haiku 5.5 (Step by Step)

1. Create or sign in to a Claude Console account at platform.claude.com and add credits on the Billing page.
2. Create an API key and store it as the ANTHROPIC_API_KEY environment variable.
3. Install or update the SDK: pip install -U anthropic for Python or npm install @anthropic-ai/sdk for Node.
4. Set model="claude-haiku-5-5" in your request.
5. Remove temperature, top_p, top_k, assistant prefill and budget_tokens if you are porting older code.
6. Choose an effort level with output_config={"effort": "low"} (or medium, high, xhigh, max). Leave it out to use medium.
7. Give max_tokens enough room, because thinking tokens count toward it.
8. Read the response by block type, since it can start with a thinking block.
9. Add cache_control to any repeated prefix over 512 tokens.

Code Example

This Python example classifies a support ticket at low effort, which is the cheapest and fastest setting for simple, high-volume work.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=1024,
    output_config={"effort": "low"},
    system="Classify the ticket as billing, bug, feature_request or other. Reply with the label only.",
    messages=[
        {"role": "user", "content": "I was charged twice for my plan this month."}
    ],
)

# Responses can start with a thinking block, so pick text blocks by type
for block in response.content:
    if block.type == "text":
        print(block.text)

print(response.usage)

For a simple task like this you can also turn thinking off. That works at high effort or below:

response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=256,
    thinking={"type": "disabled"},
    messages=[{"role": "user", "content": "Summarize: The meeting moved to Friday."}],
)

The same request with curl:

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": 1024,
    "output_config": {"effort": "low"},
    "messages": [{"role": "user", "content": "Give me three title ideas for a cooking blog."}]
  }'

Anthropic’s effort guidance for Haiku 5.5: start at medium for most work, including agentic coding. Use low for chat and simple high-volume requests. Use high for knowledge work and strict instruction following. Use xhigh or max only when your own evals show a gain.

Rate Limits

Claude API rate limits depend on your organization’s usage tier. These are the Haiku 5.5 limits from Anthropic’s rate limits page.

Cache reads do not count toward the input tokens per minute limit. Only uncached input and cache writes count. So with an 80% cache hit rate and a 2M limit, you could push about 10M total input tokens per minute. Caching is the easiest way to raise your real throughput.

Monthly spend caps also apply per tier: $500 on Start, $1,000 on Build and $200,000 on Scale.

Migrating From Haiku 4.5

Warning: Anthropic warns that code written for Haiku 4.5 can break on Haiku 5.5. Here are the changes and fixes.

Note: Also note that thinking blocks from all earlier assistant turns now stay in context and count as input tokens. Use thinking block clearing if long chats get expensive.

My Take

Haiku 5.5 changes the default answer to “which Claude model for high-volume work”. At $0.10 input, it is cheaper than Haiku 4.5 by a wide margin while being much smarter, and its Index v5 score shows it can stand next to the best open-weight models without you running any GPUs.

The traps are the 100K price step, the bigger tokenizer and default thinking. Keep prompts under 100K, start at low or medium effort, and cache your system prompt. Do that and Haiku 5.5 is one of the best value models on the market right now.

Troubleshooting: Common Problems and Fixes

400 error about temperature or top_p. Haiku 5.5 does not accept these sampling settings. Remove them from your request.

400 error about prefill. Your last message is from the assistant. End with a user message instead.

Empty text, only a thinking block. Your max_tokens is too small, because thinking counts toward it. Raise max_tokens or lower the effort.

400 when disabling thinking. thinking: {"type": "disabled"} only works at high effort or below. At xhigh or max, use adaptive thinking or omit the field.

Bill higher than expected. Check if prompts cross 100K tokens, which moves the whole request to $0.50 / $2.50. Also count thinking tokens and the roughly 30% tokenizer increase.

Refusals on normal requests. Haiku 5.5 safety classifiers can decline a request with stop_reason: "refusal". Handle it in code and rephrase or route to another model.

429 rate limit errors. You hit your tier’s RPM or token limit. Add caching, back off and retry, or move to a higher tier. See how to fix Claude 429 and 529 errors.

Frequently Asked Questions

How much does Claude Haiku 5.5 cost?

$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens. Above 100K the whole request costs $0.50 input and $2.50 output. Batch processing halves both.

What is the Claude Haiku 5.5 context window?

1M tokens, with up to 128K output tokens. On the Message Batches API, output can reach 300K with the output-300k-2026-03-24 beta header.

Is Haiku 5.5 better than open-weight models?

On Artificial Analysis Intelligence Index v5, Haiku 5.5 scores 36, while the best downloadable open-weight models score about 32, as Business Insider reported. A Kimi model scored higher but has not released its weights.

Should I use Haiku 5.5 or Sonnet 5.5?

Use Haiku 5.5 for high-volume, fast, well-defined tasks under 100K tokens. Use Sonnet 5.5 for serious coding, complex agents, or prompts that often go above 100K tokens.

Does Haiku 5.5 support thinking?

Yes. Adaptive thinking is on by default, and you control depth with the effort parameter. Default effort is medium. You can disable thinking at high effort or below.

What is the Haiku 5.5 knowledge cutoff?

Anthropic lists June 2026 as both the reliable knowledge cutoff and the training data cutoff.

Can I use Haiku 5.5 on AWS or Google Cloud?

Yes. It is on Amazon Bedrock (anthropic.claude-haiku-5-5), Claude Platform on AWS, Google Cloud and Microsoft Foundry, as well as the Claude API.

Sources

• Claude Haiku 5.5 overview, Claude Platform Docs
• What’s new in Claude Haiku 5.5, Claude Platform Docs
• Claude API pricing, Claude Platform Docs
• Effort parameter, Claude Platform Docs
• Rate limits, Claude Platform Docs
• Claude Platform release notes (October 7, 2026 entry)
• Anthropic has released Claude Haiku 5.5, Artificial Analysis
• Artificial Analysis rankings and private tests, Business Insider
• Artificial Analysis Index v5 and open-weight models, Trending Topics