Claude Haiku 5.5: Pricing, Specs, Benchmarks and What Changed

Claude Haiku 5.5 is Anthropic's smallest and fastest current model, released on 7 October 2026. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens (as of October 2026), has a 1M-token context window and 128K output, and is the first Haiku with an adjustable effort setting. Anthropic calls it "the cheapest, fastest, and most capable small model we've ever released" and says it costs about 75% less to run than Haiku 4.5 on average. It's built for classification, extraction, routing and subagent work.

Key facts

What's new in Claude Haiku 5.5

Effort levels on a Haiku for the first time

Haiku 5.5 uses adaptive thinking: it decides when and how much to think, and you steer it with the effort parameter. Anthropic's announcement lists Low, Med, High, Xhigh and Max, with medium as the API default. You can still switch thinking off with thinking: {"type": "disabled"} at high effort or below, but Anthropic recommends using effort instead.

Bigger context and output

The window grows from 200K to 1M tokens and output from 64K to 128K. Thinking tokens count toward max_tokens, so a tight limit can end a reply after the thinking block with no text. If you've set small limits for Haiku 4.5, raise them.

Much lower price, with a tokenizer catch

Here's the pricing as published by Anthropic, per 1M tokens:

Anthropic says about 90% of Haiku 4.5 requests were under 100K tokens, where 5.5 is 90% cheaper per token. The catch: Haiku 5.5 uses the newer tokenizer from Claude 4.7 and later, so the docs say the same text counts as roughly 30% more tokens. Anthropic's "about 75% cheaper on average" figure already accounts for that. Re-run your token counts before you update a cost model.

Batch API requests are 50% off. The minimum cacheable prompt drops to 512 tokens, from 4,096 on Haiku 4.5, so shorter system prompts can now be cached.

Built to be a subagent

Anthropic positions Haiku 5.5 as a subagent alongside Opus 5.5 and Sonnet 5.5 on coding tasks, and for speed-sensitive work like live customer support and browser use. A browser use tool is new for Haiku 5.5 on the Claude API and Google Cloud. If you're designing multi-model agents, our AI agent frameworks guide and how to build an AI agent cover routing between a big and a small model.

Benchmarks (Anthropic-reported)

From Anthropic's announcement. GPT-6 Luna figures are as reported by Anthropic.

The jump from Haiku 4.5 is very large across the board, and Haiku 5.5 beats GPT-6 Luna on every row where both appear. The price comparison with Luna is close: OpenAI lists Luna at $0.10 input and $0.50 output for prompts up to 272K tokens, so the two cost the same per token on short prompts, and Luna is cheaper between 100K and 272K. See GPT-6 Sol vs Luna for OpenAI's side.

Anthropic also quotes customer results, for example HubSpot at 92.8% on its CRM suite and Box scoring 11 points above Haiku 4.5 at about half the latency. These are customer-reported, on their own evals.

How to use Claude Haiku 5.5

In the Claude apps

Anthropic's pricing page lists "Haiku" on Free, Pro, Max and Team plans, and describes Haiku 5.5 as the "Fastest, most cost-efficient model". The announcement itself doesn't spell out plan-by-plan availability.

On the Claude API

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Classify this ticket: billing, bug or feature request? ..."}
    ],
)

for block in message.content:
    if block.type == "text":
        print(block.text)

Note the loop: responses can begin with thinking blocks, so pick content by type, not by position.

Migration notes from Haiku 4.5

Anthropic's "What's new" page lists these breaking changes:

• Sampling parameters can error. Omit temperature, top_p and top_k.
• Manual extended thinking errors. Replace budget_tokens with adaptive thinking and effort.
• Assistant prefill errors. End messages with a user turn.
• Old computer-use tool version errors. Replace computer_20250124 with computer_toolset_20260801 (Claude API, Google Cloud) or computer_20251124 (Bedrock).
• No structured outputs on Amazon Bedrock. Describe the format in the prompt and validate in code.
• Handle refusals. Safety classifiers can return stop_reason: "refusal".

Two behaviour changes affect cost: thinking blocks from all earlier turns stay in context and count as input, and thinking text is omitted by default (set thinking.display to "summarized" to see it).

Where Haiku 5.5 fits

Our view: start every high-volume job on Haiku 5.5 at low or medium effort, measure, and only move up a tier where it fails. With a 1M window, also check whether context engineering can do more than a bigger model. For the wider field, see the best LLM guide.