Claude Haiku 5.5 arrives on Amazon Bedrock; Anthropic claims ~75% lower cost than Haiku 4.5

Claude Haiku 5.5 is now available on Amazon Bedrock and Claude Platform on AWS. According to Anthropic, it is the fastest and most efficient model in the Claude 5.5 family, and it costs around 75% less than Claude Haiku 4.5 for most tasks. AWS announced the availability in a post on its Machine Learning blog.

Key details

• Model ID on Bedrock: global.anthropic.claude-haiku-5-5
• Positioning: subagents and high-volume, cost-sensitive workloads
• New capability: the first Haiku model with effort controls, which let developers tune cost against intelligence per task rather than per workload
• Modalities: supports high-resolution images
• Pricing: per-token pricing not disclosed in the announcement; AWS points to its Bedrock pricing page. The ~75% reduction versus Haiku 4.5 is Anthropic's claim and applies to "most tasks."
• Context window, parameter count, training cutoff, benchmark scores: not disclosed in the announcement

What Anthropic claims about capability

According to Anthropic, Haiku 5.5 is its most capable Haiku model across coding, tool use, computer use and agentic tasks. The AWS post lists these target uses:

• Coding: routing requests, reviewing code, classifying long documents, iterating on UI and UX changes, and making small, specific changes across multiple files
• Knowledge work: extracting information from small-to-medium documents, initial scans, and quick questions over a knowledge base
• Computer use: acting as a subagent for repetitive browser and desktop tasks
• Traditional NLP: classification, summarization and text generation at production volume

The AWS sample code notes that Haiku 5.5 may return a thinking block before the text block. Developers should select the text block by type rather than by fixed index.

Opus 5.5 and Haiku 5.5 as a pair

AWS describes Haiku 5.5 as a companion to the recently announced Claude Opus 5.5. In the suggested pattern, Opus 5.5 plans the work and handles the hardest reasoning, such as release debugging, security review of large pull requests and long analyses. Haiku 5.5 runs the fast subagent layer: routing, classifying, summarizing, rewriting long documents and applying small edits across many files. Haiku can also act as a review subagent that checks order of operations and high-level direction, so Opus spends tokens on the hardest problems. AWS notes that Haiku's speed and cost make it practical to run many subagents in parallel.

Access and availability

On Amazon Bedrock, Haiku 5.5 is available through these inference profiles on bedrock-runtime:

• US Geo CRIS (us.)
• EU Geo CRIS (eu.)
• AU Geo CRIS (au.)
• JP Geo CRIS (jp.)
• Global CRIS (global.)

In AWS GovCloud (US), it is available on both the bedrock-runtime and bedrock-mantle endpoints. It is also available through Claude Platform on AWS in North America.

Developers can call it through the Anthropic Messages API via the Anthropic SDK, or through the InvokeModel and Converse APIs using the AWS CLI and SDKs. On Bedrock, data stays within AWS infrastructure with Regional data residency. Access works with IAM, CloudTrail, CloudWatch and Bedrock Guardrails, and usage appears on the AWS bill. Claude Platform on AWS offers Anthropic's native platform experience through the AWS Management Console, with AWS billing and authentication.

What this means

The most consequential change is effort controls on a Haiku-tier model. Until now, cost-versus-quality tuning in a multi-agent pipeline mostly meant choosing between model tiers. Per-request effort settings let teams use one cheap model for both trivial routing and moderately demanding subtasks.

The ~75% cost reduction is the number to verify. It is Anthropic's claim, and "for most tasks" leaves room for variation, partly because effort settings and thinking blocks can change token consumption. Teams should benchmark on their own workloads once Bedrock pricing is checked.

The missing context window and benchmark data limit direct comparison with rival small models for now. The Opus-plans, Haiku-executes pattern still shows where Anthropic expects Haiku to be used: as the high-volume execution layer in agent systems, with a large model reserved for planning and judgment.