Claude Haiku 5.5

Summary

Anthropic released Claude Haiku 5.5 on 2026-10-07, completing the Claude 5.5 family that Opus 5.5 (2026-09-22) and Sonnet 5.5 (2026-09-28) started. It is the smallest and fastest model in the lineup, aimed at high-volume, cost-sensitive work such as classification, extraction, and routing. It is the first Haiku with an adjustable effort setting, and its price depends on prompt length: prompts over 100,000 tokens cost five times as much.

Key points

API ID claude-haiku-5-5; on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
Prices per 1M tokens (as of 2026-10-08): $0.10 input / $0.50 output for prompts up to 100K tokens; $0.50 / $2.50 above that. Cache reads $0.01 (≤100K) or $0.05 (>100K). Haiku 4.5 costs $1 / $5, so short prompts are 10x cheaper than on the previous Haiku.
1M-token context window, 128K max output, reliable knowledge cutoff June 2026 (models overview, checked 2026-10-08). Retirement not before 2027-10-07.
Adaptive thinking with effort levels low to max; default effort medium.
Uses the newer tokenizer of the Claude 4.7+ generation. Simon Willison measured about 1.25x more tokens than Haiku 4.5 for the same text.
Supports computer use and browser use.
Anthropic's reported scores: OSWorld 2.1 (offline subset) 72.4%, GDPval-AA v2.1 1,620 Elo, Humanity's Last Exam 45.9% (no tools) / 57.4% (with tools), Terminal-Bench 4.0 39.2%, FrontierCode 1.1 46.4%. The page as fetched also lists Chartography at 46.4%; the identical figure may be a page or extraction error.
Claude Code 2.1.293 made Haiku 5.5 its default Haiku model.

Why it matters

Haiku 5.5 brings the 1M context window, effort control, and computer use of the larger 5.5 models down to a price below most small models. The tiered price is new for Claude: until now Claude 4.6+ models charged the same rate across the whole 1M window. Cost estimates for long-document work on Haiku 5.5 must account for the higher tier.