On October 7, 2026, Anthropic launched Claude Haiku 5.5 with pricing that cuts token costs by up to 90% versus its predecessor. The model targets high-throughput workloads — agents, classifiers, document processing — where per-call economics matter more than raw capability.
The Numbers
For prompts up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. That compares to Haiku 4.5's rates of $1.00 and $5.00 — a 90% reduction at the short-context tier.
The pricing splits at the 100K-token mark. Longer contexts move to a separate tier: $0.50 per million input tokens and $2.50 per million output tokens. So the headline "90% cheaper" applies most strongly to the typical chatbot or agent turn, not to deep-context tasks like full codebase analysis or very long document review.
Batch and cache rates round out the picture: Cache Read at $0.01/M tokens, Cache Write at $0.125/M tokens, and batch rates of $0.05/$0.25 per million (input/output) for short-context requests. Anthropic estimates that, across a typical mix of real workloads, the average cost reduction lands around 75%.
What You Actually Get
Haiku 5.5 is Anthropic's fastest and most affordable model in the Claude 5 family. It isn't intended to replace Sonnet or Opus for complex reasoning — it's built for tasks where you need thousands of calls per day to be economically viable. Think document routing, intent classification, summarization pipelines, or the lightweight decision steps inside a larger multi-agent system.
At $0.10 per million input tokens, a million short agent turns — each consuming a few hundred tokens — costs roughly $50. That's a different economic conversation than the $500 the same volume would have cost on Haiku 4.5.
Why This Matters for AI Applications
The cost of inference has been dropping steadily, but the pace accelerated sharply with the Claude 5 generation. Haiku 5.5's pricing brings capable AI into use cases that were previously too expensive to justify at scale — particularly in enterprise settings where CFOs were keeping a close eye on token bills.
For developers building on the Anthropic API, this is a direct invitation to revisit architectures that were previously cost-constrained. Workflows that relied on simpler models or rule-based systems to keep costs down may now make more sense rebuilt with a capable but cheap model at their core.