Anthropic has released Claude Haiku 5.5, which the company calls the cheapest, fastest, and most capable small model it has ever shipped. The model is aimed squarely at developers running high-volume, cost-sensitive workloads, and it arrives alongside price cuts for Claude Sonnet 5.5 and a new API credit program for subscribers. For teams building AI agents and applications, it signals that Anthropic’s strategy increasingly hinges on layering cheap, fast models underneath its flagship ones rather than relying on a single do-everything system.

What Claude Haiku 5.5 actually does

Haiku is Anthropic’s “small model” tier, sitting below Sonnet and Opus in the Claude lineup. Historically, small models trade some reasoning depth for speed and lower cost, making them well suited to tasks that are repetitive, high-volume, or latency-sensitive. Haiku 5.5 continues that pattern: Anthropic positions it for summarization, context compaction, database queries, and classification requests, as well as live customer support and browser automation where response speed matters as much as accuracy.

A key architectural detail is that Haiku 5.5 is Anthropic’s first Haiku-class model with an adjustable effort setting. This lets developers dial the model toward either lower cost or higher intelligence depending on the task, a flexibility previously reserved for Anthropic’s larger models. Anthropic illustrates this trade-off using three benchmarks: OSWorld 2.1 (how well an agent can operate a real computer through long, multi-step tasks), GDPval-AA v2.1 from Artificial Analysis (real-world professional work across 44 occupations), and Humanity’s Last Exam (expert-level academic reasoning). Performance on each scales with the effort setting chosen.

How it fits alongside Sonnet and Opus

Anthropic explicitly frames Haiku 5.5 as a companion to its bigger models rather than a replacement. It “pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work,” meaning a larger model can handle complex reasoning or planning while Haiku 5.5 is dispatched to execute narrower, well-defined subtasks quickly and cheaply. This subagent pattern is becoming common in agentic AI systems: a capable but expensive model orchestrates, while cheaper models do the grunt work at scale.

This release also comes with broader pricing changes across the Claude lineup. Haiku 5.5 itself costs around 75% less to run than Haiku 4.5 on average. Separately, Anthropic is halving the price of cache reads for Claude Sonnet 5.5, which brings Sonnet’s typical agentic-work cost down by about 20%. Anthropic is also introducing a new monthly API credit for Claude Max and Team subscribers, intended to help them experiment with building new agents and applications on the Claude Platform without separate pay-as-you-go spend.

Why early adopters are interested

Anthropic shared feedback from companies that tested Haiku 5.5 before launch, and the common thread is speed and cost efficiency at roughly equal or better accuracy than Haiku 4.5. One customer building an AI agent product for bug triage and project setup reported over a 30% reduction in task-completion latency and up to 2.5x faster inference per agent turn compared to the model they used previously. HubSpot said Haiku 5.5 scored 92.8% on its CRM evaluation suite for smaller models, the best result it has seen, and was also the fastest and most accurate on a stale-record identification task.

Other reported results include a statistically significant accuracy improvement on a document question-answering workload running roughly 8 million calls a week in production (0.84 versus 0.76 for Haiku 4.5), and an enterprise content platform scoring 11 points higher than Haiku 4.5 at about half the latency. Another tester described using Haiku 5.5 as a subagent that pulls specific data, like a segment revenue line from a 10-K filing, while a larger model builds the overall output, citing it as “accurate enough” to trust for that narrow role while being cheap enough to run frequently.

What this means for developers in practice

For teams already using Claude models, Haiku 5.5 is most relevant where cost scales with volume: customer support chat, classification pipelines, retrieval-augmented question answering, and agent subagents that handle repetitive lookups. The adjustable effort setting means developers don’t have to choose a single fixed trade-off; they can tune a given deployment toward speed and cost or toward intelligence depending on the task’s tolerance for error.

The combination of a cheaper Haiku, a cheaper Sonnet, and new API credits also lowers the barrier for experimenting with multi-model agent architectures, where a small model like Haiku handles volume and a larger model like Sonnet or Opus handles judgment-heavy steps. The main caveat is that all performance figures and customer testimonials above come from Anthropic’s own benchmarks and early-access customers, so independent, head-to-head comparisons against competing small models from other vendors are not yet part of the public record.

Frequently asked questions

Is Claude Haiku 5.5 meant to replace Sonnet or Opus? No. Anthropic positions it as a lightweight complement, best used for high-volume or latency-sensitive tasks, or as a subagent supporting larger models rather than as a standalone replacement for complex reasoning work.

How much cheaper is it than the previous Haiku model? Anthropic states that Haiku 5.5 costs around 75% less on average to run than Haiku 4.5.

What is the “effort setting” on Haiku 5.5? It is a configurable option that lets developers choose whether the model should optimize for lower cost and speed or for higher intelligence and accuracy, depending on the benchmark and task requirements.

Source: Claude Haiku 5.5