Claude Haiku 5.5 reaches Kiro in GovCloud at 0.06x credit cost

Amazon Web Services has made Anthropic’s Claude Haiku 5.5 available in the Kiro IDE and CLI for the AWS GovCloud (US) Regions, according to its announcement of November 2. The model runs at a 0.06x credit multiplier.

With the addition, all three Claude 5.5 models (Haiku, Sonnet and Opus) can now be selected in Kiro inside GovCloud, the isolated AWS regions used for US government and other regulated workloads.

A 1M context window at a fraction of the credit cost

AWS describes Haiku 5.5 as the fastest model in the Claude 5.5 family and a clear step up from Claude Haiku 4.5. It targets high-volume jobs where response time matters: condensing long conversations or documents, sorting incoming requests and sending them to the right place, running text through bulk pipelines, and making narrow edits in many files at once.

The model uses adaptive thinking with five effort levels, from low to max, and medium is the default. It comes with a 1M context window. Billing is by Kiro credits: the multiplier is 0.06x for normal requests, and requests above 100K tokens are billed at 0.3x. That is five times the base rate, so a long-context job costs noticeably more per request than a short one, though the page does not say what a credit costs in dollars or how the multipliers compare with those of Sonnet 5.5 and Opus 5.5.

To use it, customers need the newest release of the IDE or CLI, which must be restarted before Haiku 5.5 appears in the model selector. AWS points customers to the GovCloud documentation, a monitoring and tracking guide, and their account team for details.

Cheap workers beside the expensive planners

The most telling line in the announcement is about division of labor. AWS suggests running Haiku 5.5 as a parallel reviewer, leaving the tougher planning and coding to the two larger models, Opus 5.5 and Sonnet 5.5. In practice that is a routing pattern: a large model decides what to build, and many small, cheap calls check, summarize and patch around it.

The pricing is built for that. A multiplier of 0.06x means a team can fan out many parallel reviewers without the bill growing at the rate it would with a larger model. The threshold at 100K tokens, however, is a reminder that feeding a full 1M window into a cheap model is not free: AWS charges five times more once a request passes that size.

The release also fits a run of GovCloud additions for AWS. On October 30 the company said web search had arrived in Kiro in GovCloud (US), giving the agent real-time access to current information with citations. In the same post, AWS said enterprise administrators can switch off web search and web fetch for their organization from the Kiro console, and that teams can strip web tools from individual custom agents. Also on October 30, AWS announced expanded API support and Cross Region Inferencing for OpenAI models on Amazon Bedrock in GovCloud, and our own report covered how the ChatGPT desktop app can now run on Bedrock inside AWS GovCloud.

Taken together, the pattern is that regulated customers are getting current-generation models and agent features within days of each other rather than months behind commercial regions, though the sources do not state a general timetable.

What the announcement leaves open

The performance claims are AWS’s own: the page gives no benchmark scores, latency figures or speed comparison with Haiku 4.5, so “fastest” and “a clear step up” are not independently verified. It also does not say which specific GovCloud Regions are covered beyond the “(US)” label, whether the model is available through Amazon Bedrock there, or what the credit cost is in dollars under each Kiro plan.

For developers already on Kiro in GovCloud, the practical change is simple: a cheaper, faster option now sits in the model selector next to Sonnet 5.5 and Opus 5.5. Whether it saves money in a real workflow will depend on how much work can be pushed to the small model and how often requests cross the 100K-token line.