Claude Haiku 5.5: EU pricing and setup 2026

TL;DR

Claude Haiku 5.5 is the fastest current model for classification, extraction and routing. It costs 0.10 / 0.50 USD per million tokens up to 100,000 prompt tokens, 0.11 / 0.55 on EU endpoints. In the EU pin Bedrock eu.anthropic.claude-haiku-5-5 or Google Europe multi-region. Foundry has no EU data zone. New pipelines should not stay on Haiku 4.5.

What is Claude Haiku 5.5 and when is it the right model?

Claude Haiku 5.5 is the production model for high volume and low latency. Anthropic names classification, routing, extraction and subagent work as the core jobs. It shipped on 7 October 2026, with a 1 million token context and at most 128,000 output tokens. Reliable knowledge ends in June 2026. New pipelines pin claude-haiku-5-5, not Haiku 4.5.

The Haiku 5.5 overview and the models overview from 23 October 2026 put Haiku as the fastest model in the current lineup. Sonnet 5.5 stays the speed and intelligence balance. Opus 5.5 stays the long-agent line. The cost split between those two IDs is in Claude Sonnet vs Opus. Token rates for every model sit in Claude API pricing in the EU.

Feature	Claude Haiku 5.5	Claude Haiku 4.5	Claude Sonnet 5.5
Status	Active (latest)	Active (legacy)	Active (latest)
Release	7 October 2026	2025	28 September 2026
Retirement (Anthropic)	No sooner than 7 October 2027	No sooner than 15 October 2026	No sooner than 28 September 2027
Claude API ID	claude-haiku-5-5	claude-haiku-4-5	claude-sonnet-5-5
List price input / output	0.10 / 0.50 USD (up to 100k)	1 / 5 USD	2 / 10 USD
Context / max output	1M / 128K	200K / 64K	1M / 128K
Default effort	medium	none	high
Thinking	Adaptive, can turn off	Manual budget_tokens	Adaptive

One million tokens is about 555,000 words on the tokenizer introduced with Claude Opus 4.7. The same text counts as about 30 percent more tokens on Haiku 5.5 than on Haiku 4.5. A budget built on 4.5 counts will run short. Adaptive thinking is on. Steer depth with effort: low, medium, high, xhigh, max. Omit temperature, top_p and top_k or the request returns 400.

We recommend Haiku 5.5 as the default for ticket routing, PII flags, JSON extraction and as a subagent in front of a Sonnet pass. For code review, long agents and work that needs a costly reasoning path, stay on Sonnet 5.5. IDs and the EU call for Sonnet are in Claude Sonnet 5 in Europe.

What does Claude Haiku 5.5 cost in the EU?

Claude Haiku 5.5 costs 0.10 USD input and 0.50 USD output per million tokens on the Claude API while the prompt stays at or under 100,000 tokens. Above that line the row jumps to 0.50 / 2.50. Regional and multi-region endpoints add 10 percent from Haiku 4.5 onward. Google lists Europe at 0.11 / 0.55. No vendor publishes EUR list prices.

The figures are from 23 October 2026 on the Anthropic pricing page and the Google Cloud partner table. Anthropic applies the 10 percent rule from Sonnet 4.5, Haiku 4.5 and Opus 4.5 on regional Bedrock endpoints and Google multi-regions. The Bedrock pricing page did not print a dollar table for Haiku 5.5 in this run. Budget EU work on the Anthropic row plus 10 percent until Cost Explorer shows the Marketplace rate. The Marketplace product ID on the Bedrock card is prod-6cyn7tgqazjhu.

Route	Input / output up to 100k	Input / output over 100k	5-min write / read (up to 100k)	Batch in / out (up to 100k)
Claude API global	0.10 / 0.50	0.50 / 2.50	0.125 / 0.01	0.05 / 0.25
Google Cloud, first table	0.10 / 0.50	0.50 / 2.50	0.125 / 0.01	0.05 / 0.25
Google Cloud Europe	0.11 / 0.55	0.55 / 2.75	0.1375 / 0.011	0.055 / 0.275
Bedrock EU geo, 10% rule	0.11 / 0.55	0.55 / 2.75	0.1375 / 0.011	50% of the EU row if batch is enabled
US-only, Claude API or Foundry US zone	1.1x the global row	1.1x	1.1x	1.1x

USD per million tokens, standard, as of 23 October 2026. Foundry and Claude Platform on AWS bill in Claude Consumption Units at 0.01 USD per CCU. One hundred CCU equal 1.00 USD of list fees. That is a billing unit, not a discount. The 1.1x rule for inference_geo: "us" also applies to the higher long-context row.

Worked example, our arithmetic, not a vendor quote: 10,000 classifications with 800 input and 80 output tokens cost 1.20 USD globally on Haiku 5.5. The same volume on the Europe row costs 1.32 USD. The same volume on Sonnet 5.5 Europe costs 26.40 USD. Batch on the Europe Haiku row costs 0.66 USD.

Prompt caching pays back after one 5-minute cache read, or after two 1-hour reads. The minimum cacheable prompt is 512 tokens, down from 4,096 on Haiku 4.5. Breakpoints, TTL and the EU call are in Claude prompt caching in the EU.

US-only inference on the Claude API (inference_geo: "us") costs 1.1x. That is not an EU product. The data residency page accepts only global and us. Teams that need processing in EU member states pay the 10 percent on Bedrock or Google Cloud, not on the Claude API. The four-route residency matrix is in Claude data residency in the EU.

Which model IDs apply in the EU?

For Claude Haiku 5.5 pin eu.anthropic.claude-haiku-5-5 on Bedrock and claude-haiku-5-5 with region="eu" on Google Cloud. The bare Bedrock ID anthropic.claude-haiku-5-5 is the card ID, not the on-demand modelId on bedrock-runtime. Foundry and the Claude API use claude-haiku-5-5 with no EU prefix. The Claude API has no EU inference.

The Bedrock card lists geo IDs us., eu., au. and jp. plus global.anthropic.claude-haiku-5-5. From Frankfurt (eu-central-1) the destinations are Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. London as a source adds London as a destination. Zurich adds Zurich. Both sit outside the EU. Call the profile from a region in an EU member state, typically eu-central-1. In-region on bedrock-runtime is N/A on the card. The global sample from us-east-1 with global.anthropic.claude-haiku-5-5 leaves the EU.

Google lists on the Haiku 5.5 page model availability in the Europe multi-region, the US multi-region and on the global endpoint. ML processing names Europe, the United States and asia-southeast1. The multi-region quota is 1,500 requests per minute, 15 million input tokens per minute (uncached plus cache write) and 1.5 million output tokens per minute. The global endpoint doubles that (3,000 QPM, 30 million input TPM, 3 million output TPM). Context length is 1,000,000 tokens. Set region="eu". The residency note excludes the United Kingdom and Switzerland from the Europe endpoint.

Foundry is the outlier. Microsoft’s Claude models page places Haiku 5.5 on Hosted on Azure (Global Standard and Data Zone Standard US) and on Hosted on Anthropic (Global Standard). There is no Europe data zone. The quota page lists pay-as-you-go at 40 RPM, 40,000 ITPM and 8,000 OTPM, and Enterprise or MCA-E at 10,000 RPM, 10 million ITPM and 2 million OTPM. Cache reads do not count toward ITPM. The four-provider matrix is in Claude GDPR in Europe.

How do you call Claude Haiku 5.5 in the EU?

You need a cloud account in an EU region, the model ID with an EU prefix or region="eu", and you must not copy the global AWS sample. On Bedrock the Frankfurt call is eu.anthropic.claude-haiku-5-5. On Google Cloud the ID stays claude-haiku-5-5 and the region is eu. Five steps cover the first production call.

1. Create the AWS account or Google Cloud project in a region of an EU member state. For Bedrock, eu-central-1 (Frankfurt) is the usual source region.
2. Enable the model. On Bedrock use Model Access or AWS Marketplace (product ID prod-6cyn7tgqazjhu). On Google Cloud use Model Garden, model claude-haiku-5-5.
3. Pin the EU ID in code. No generic haiku, no global. prefix, no us-east-1.
4. Send a test prompt and check the metadata. On Bedrock the modelId must start with eu.. On Google Cloud the region must be eu.
5. Set a cap: a budget alarm, a LiteLLM spend cap, or both. Token prices alone will not stop a pipeline.

import boto3

client = boto3.client("bedrock-runtime", region_name="eu-central-1")
response = client.converse(
    modelId="eu.anthropic.claude-haiku-5-5",
    messages=[{"role": "user", "content": [{"text": "Classify in one word: late delivery."}]}],
)
print(response["output"]["message"]["content"][0]["text"])

from anthropic import AnthropicVertex

client = AnthropicVertex(region="eu", project_id="YOUR_PROJECT")
message = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=256,
    messages=[{"role": "user", "content": "Classify in one word: late delivery."}],
)
print(message.content[0].text)

Read response blocks by type, not by position. Adaptive thinking can place a thinking block before any text. A small max_tokens then stops after thinking and before text. low keeps classification cheap. medium is the default. Bedrock basics, and why the global sample leaves the EU, are in Amazon Bedrock in Europe.

What breaks when you move from Haiku 4.5 to Haiku 5.5?

Haiku 4.5 code breaks on Haiku 5.5 in five places. budget_tokens returns 400. temperature, top_p and top_k return 400. A prefill at the end of messages returns 400. The computer-use tool computer_20250124 returns 400. The same text counts as about 30 percent more tokens. Treat the move as a checklist, not an ID swap.

The migration guide and What’s new list the breaks. Replace thinking: {"type": "enabled", "budget_tokens": N} with {"type": "adaptive"} and steer with output_config.effort. You can still turn thinking off with thinking: {"type": "disabled"} at high effort or below. Anthropic still prefers the effort lever. Summarized thinking arrives only if you set thinking.display to "summarized".

On the Claude API and Google Cloud (Google managed), computer use belongs in computer_toolset_20260801. On Bedrock use computer_20251124 with the computer-use-2025-11-24 header. Browser use (browser_toolset_20260801) exists on the Claude API and Google, not on Haiku 4.5. Priority Tier is not supported on Haiku 5.5. A 2,000 by 1,500 pixel image costs about 2.5 times as many visual tokens as on 4.5, because Haiku 5.5 uses the high-resolution tier (downscale above 2,576 pixels on the long edge or 4,784 visual tokens).

Older thinking blocks stay in context and count as input. Haiku 4.5 kept only the last turn. Multi-turn chats get more expensive than the tokenizer change alone. Thinking blocks work only in the account that produced them, or in a linked account. You must handle stop_reason: "refusal" yourself. There is no server-side fallback. The 50 percent batch discount and 300k output with header output-300k-2026-03-24 belong to the Message Batches API, not automatically to every cloud.

FAQ

How much does Claude Haiku 5.5 cost in the EU per million tokens?

0.11 USD input and 0.55 USD output while the prompt stays at or under 100,000 tokens. Above that, 0.55 / 2.75. That is the 10 percent rule on the global row 0.10 / 0.50 or 0.50 / 2.50. Google lists the same Europe figures. Batch halves input and output. Cache reads cost 0.011 USD per million tokens under the 100k line.

Claude Haiku 5.5 vs Claude Sonnet 5.5: which model for production?

Haiku 5.5 for high volume with a short answer: routing, extraction, a pre-filter. Sonnet 5.5 for knowledge, code and agents where a miss costs more than tokens. Both have a 1 million token context and Bedrock EU IDs. Sonnet’s list price is twenty times Haiku while the prompt stays under 100,000 tokens.

Which Bedrock ID applies in Frankfurt?

eu.anthropic.claude-haiku-5-5 on bedrock-runtime in eu-central-1. The bare ID anthropic.claude-haiku-5-5 is the card ID, not the on-demand modelId. In-region on that endpoint is N/A. The global sample from us-east-1 with global.anthropic.claude-haiku-5-5 leaves the EU.

Does Claude Haiku 5.5 on Microsoft Foundry stay in the EU?

No. Foundry hosts Haiku 5.5 as Hosted on Azure (Global Standard and US Data Zone) and as Hosted on Anthropic (Global Standard only). There is no Europe data zone for Claude. If EU inference is mandatory, use Bedrock or Google Cloud.

Is Claude Haiku 4.5 enough instead of Claude Haiku 5.5?

For existing code, often still. For new pipelines, no. Haiku 4.5 costs 1 / 5 USD, has a 200,000 token context and Anthropic holds it no sooner than 15 October 2026. Haiku 5.5 is ten times cheaper under 100,000 tokens, has a 1 million context and a June 2026 knowledge cutoff. The move needs a checklist, not only a new ID.

Is the Claude API still an option for EU companies?

As a chat and admin product, yes. As an inference location, no. inference_geo accepts only global and us. Workspace geo is us only today. For processing in EU member states, the Claude API is the wrong endpoint. Use it for Claude.ai seats, not for a production API that carries personal data.

Sources

1. Anthropic: Claude Haiku 5.5 (23 October 2026)
2. Anthropic: What's new in Claude Haiku 5.5 (23 October 2026)
3. Anthropic: Haiku 5.5 migration guide (23 October 2026)
4. Anthropic: Pricing (23 October 2026)
5. Anthropic: Models overview (23 October 2026)
6. Anthropic: Data residency (23 October 2026)
7. AWS Bedrock: Claude Haiku 5.5 model card (23 October 2026)
8. Google Cloud: Claude Haiku 5.5 (23 October 2026)
9. Google Cloud: Agent Platform pricing (23 October 2026)
10. Google Cloud: Data residency (23 October 2026)
11. Microsoft Learn: Claude models in Foundry (23 October 2026)
12. Microsoft Learn: Claude quotas and rate limits (23 October 2026)