Anthropic launches Claude Haiku 5.5 with big price cut
Anthropic releases Claude Haiku 5.5 with a 1M token window, 90% price cut on common API calls, and native subagent support.
Compact model features a 1M token context window, up to 90% lower rates, and a sharp focus on subagent operating costs.
On October 7, Anthropic released Claude Haiku 5.5, the fastest and most affordable entry in its model lineup. The launch went live simultaneously across Anthropic’s platform API, Claude.ai, Claude Code, and partner cloud services on AWS Bedrock, Google Cloud, and Microsoft Azure Foundry. The aggressive pricing targets OpenAI directly by matching GPT-6 Luna rates at $0.10 per million input tokens and $0.50 per million output tokens.
Price war and subagent economics
The new pricing baseline covers requests up to 100,000 tokens, a band representing nearly 90% of typical Haiku query volume. In this bracket, input costs fall by up to 90% compared to Haiku 4.5, which Anthropic calculates will save developers an average of 75% on operating bills. For larger calls beyond 100,000 tokens, prices step up to $0.50 for input and $2.50 for output per million tokens. Prompt cache reads dropped to $0.01 per million.
The pricing strategy tracks how engineering teams build automated systems today. Few developers run expensive frontier models for basic data ingestion or quick verification steps. Lightweight models have become the practical workhorses of subagent pipelines, tackling text triage, short summaries, and repetitive tool calls. By matching competitor prices, Anthropic aims to stop developers from offloading routine background workloads to rival providers.
Claude Haiku 5.5 specs and 1M token context
Under the hood, Claude Haiku 5.5 provides a 1-million-token context window. Maximum output reaches 128,000 tokens per request, extending to 300,000 tokens through the Message Batches API using a beta header. The underlying training data is current through June 2026, running on the modern Claude 5 tokenizer.
Anthropic also added native support for its adaptive thinking parameter (effort), which lets developers throttle reasoning depth to fit the job. Simple classification tasks can execute without thinking delays, while tricky multi-step workflows can spend extra internal processing steps before generating an answer.
Computer use and reported benchmarks
Anthropic explicitly geared the model toward desktop and web automation (computer use and browser use), launching beta tooling in its official Python and TypeScript SDKs. The AWS Bedrock model card documents native integration for these agentic loops.
In test results published by Anthropic, Haiku 5.5 scored 72.4% on the OSWorld 2.1 desktop automation suite, compared to 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On the multidisciplinary Humanity’s Last Exam, it notched 45.9% alone and 57.4% when granted access to external tools. On the agentic terminal benchmark Terminal-Bench 4.0, the model logged 39.2%. These figures come directly from vendor lab tests and still lack independent verification under real production traffic.
Sonnet cache cuts and monthly credits
Alongside the release, Anthropic cut the cost of prompt cache reads on Claude Sonnet 5.5 by half, dropping from $0.20 to $0.10 per million tokens. That discount shaves around 20% off operational expenses for architectures that repeatedly reference long system instructions or codebase prefixes.
Enterprise customers and paid tiers also received monthly API credits between $100 and $500. Anthropic is pairing the two tiers into a combined pipeline strategy: Haiku runs high-volume, cheap subagent routines, while a discounted Sonnet handles orchestrator decisions.