Anthropic Launches Claude Haiku 5.5, Its Fastest and Cheapest Model Yet

Anthropic launched Claude Haiku 5.5, positioning it as its fastest, cheapest, and most capable small model to date, built for high-volume, cost-sensitive work such as summarization, data compaction, classification, and acting as a subagent alongside larger Claude models.

Details

• Pricing: for requests up to 100k tokens, input is $0.10/million tokens and output $0.50/million — about 90% cheaper than Haiku 4.5 for the roughly 90% of typical usage that falls under 100k tokens; above 100k tokens, pricing is $0.50/$2.50 per million (input/output), roughly 50% cheaper than before. Cache reads cost $0.01–0.05/million and cache writes $0.125–0.625/million
• Benchmarks: scored 1,620 Elo on GDPval-AA v2.1 knowledge-work evaluation (vs. 735 for Haiku 4.5), 72.4% on OSWorld 2.1 computer-use tasks, 57.4% on Humanity’s Last Exam with tools, 39.2% on Terminal-Bench 4.0 agentic coding, and 46.4% on the Chartography visual-reasoning benchmark without tools
• Adjustable effort settings: Low, Medium, High, Xhigh, and Max modes let developers trade off cost against intelligence per request
• Customer results: Asana reported over 30% lower latency and up to 2.5x faster inference; HubSpot hit 92.8% on its CRM evaluation suite, beating all competitors tested; AlphaSense saw a statistically significant accuracy jump (0.84 vs. 0.76) across 8 million weekly production queries; Box scored 11 points higher than Haiku 4.5 at roughly half the latency; Cognition’s Devin/Fusion reached a 66.2 FrontierCode score using Haiku 5.5 as a coding sidekick
• Safety: alignment evaluations show fewer misaligned behaviors and reduced willingness to assist with misuse versus Haiku 4.5; cybersecurity safeguards are tighter than Haiku 4.5 but allow more defensive use than Sonnet 5.5; biology safeguards match the Sonnet and Opus lines
• Availability: live immediately on the Claude Platform (model ID claude-haiku-5-5), AWS, Google Cloud, and Microsoft Azure
• Related updates: Sonnet 5.5 cache-read pricing cut 50% to $0.10/million tokens (roughly 20% lower typical agentic workload cost); Max and Team subscribers now get $100–500 in monthly API credits; Python and TypeScript SDKs gained computer-use and browser-use (beta) support

What happened next

Haiku 5.5 extends Anthropic’s small-model line past the point where “cheap” and “capable” trade off against each other, with early enterprise adopters reporting both lower latency and higher accuracy than the previous generation — a combination that, paired with the Sonnet 5.5 cache-price cut and new API credits, signals Anthropic is pushing hard to make its stack the economical default for high-volume agentic and subagent workloads.