Claude Haiku 5.5: 75% Cheaper, Built to Be the Subagent
Anthropic released Claude Haiku 5.5 on Wednesday and Hacker News gave it 516 points in a few hours. The pitch is not intelligence. It is price and a specific role. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100k tokens, which is 90% below Haiku 4.5 on those requests and, Anthropic says, around 75% cheaper on average. Cache reads are a cent per million. The model is positioned as the subagent that runs under Opus 5.5 and Sonnet 5.5 on coding work, and as the thing you point at compaction, summaries, classification, database queries, live support and browser use.
The numbers are good for a small model. Terminal-Bench 4.0 at 39.2% against Haiku 4.5's 0.0% and GPT-6 Luna's 16.4%. OSWorld 2.1 offline at 72.4% against Luna's 48.9%. Humanity's Last Exam at 45.9% without tools. GDPval-AA v2.1 Elo of 1620, which is above Luna's 1437. FrontierCode 1.1 at 46.4%. Haiku 5.5 is the first Haiku with an adjustable effort setting, and Anthropic's own charts show Sonnet 5.5 still winning on the hard agentic tier, so the honest read is "cheap model that handles the narrow jobs well," not a frontier model at a discount. HubSpot reported 92.8% on its CRM suite, the best score it has seen from a small model. Asana measured 30% lower latency and up to 2.5x faster inference per agent turn.
Two things in the fine print matter more to agent builders than the benchmarks. Sonnet 5.5 cache reads were cut from $0.20 to $0.10 per million, which Anthropic says makes Sonnet about 20% cheaper on most agentic tasks because cache reads are most of an agent's token volume. And Max 5x subscribers now get $100 of monthly API credit, Max 20x gets $200, Team gets up to $500 pooled. That is a direct nudge: stop treating the subscription as a chat plan and build something that calls the API. The Python and TypeScript SDKs also gained beta support for computer use and browser use.
The cybersecurity safeguards are looser than Sonnet 5.5's but still block penetration testing, which is the same split Mistral called out a day earlier with Large 4. Available now on the Claude Platform, AWS, Google Cloud and Azure as claude-haiku-5-5.
Link: anthropic.com/claude-haiku-5-5
← Back to all articles
The numbers are good for a small model. Terminal-Bench 4.0 at 39.2% against Haiku 4.5's 0.0% and GPT-6 Luna's 16.4%. OSWorld 2.1 offline at 72.4% against Luna's 48.9%. Humanity's Last Exam at 45.9% without tools. GDPval-AA v2.1 Elo of 1620, which is above Luna's 1437. FrontierCode 1.1 at 46.4%. Haiku 5.5 is the first Haiku with an adjustable effort setting, and Anthropic's own charts show Sonnet 5.5 still winning on the hard agentic tier, so the honest read is "cheap model that handles the narrow jobs well," not a frontier model at a discount. HubSpot reported 92.8% on its CRM suite, the best score it has seen from a small model. Asana measured 30% lower latency and up to 2.5x faster inference per agent turn.
Two things in the fine print matter more to agent builders than the benchmarks. Sonnet 5.5 cache reads were cut from $0.20 to $0.10 per million, which Anthropic says makes Sonnet about 20% cheaper on most agentic tasks because cache reads are most of an agent's token volume. And Max 5x subscribers now get $100 of monthly API credit, Max 20x gets $200, Team gets up to $500 pooled. That is a direct nudge: stop treating the subscription as a chat plan and build something that calls the API. The Python and TypeScript SDKs also gained beta support for computer use and browser use.
The cybersecurity safeguards are looser than Sonnet 5.5's but still block penetration testing, which is the same split Mistral called out a day earlier with Large 4. Available now on the Claude Platform, AWS, Google Cloud and Azure as claude-haiku-5-5.
Link: anthropic.com/claude-haiku-5-5
Comments