Anthropic releases cheaper Claude Haiku 5.5 with new safety scores

Anthropic released Claude Haiku 5.5 on October 7, 2026, at $0.10/$0.50 for short prompts, about 75% below Haiku 4.5, with stronger safety results.

Anthropic released Claude Haiku 5.5 on October 7, 2026, its fastest small model for high-volume work such as summaries, classification, support, and coding subagents, available as claude-haiku-5-5 on the Claude Platform, AWS, Google Cloud, and Microsoft Azure; MarkTechPost and Latent Space say it has a 1 million-token context, and Latent Space says Haiku 4.5 had 200K. For prompts up to 100,000 tokens, input is $0.10 and output $0.50 per million tokens with cache reads at $0.01; Anthropic and several outlets say it costs about 75% less to run than Haiku 4.5, while The Decoder and Help Net Security say short-prompt list prices are up to 90% lower than Haiku 4.5's $1 input and $5 output, and Simon Willison and The Decoder say a new, less generous tokenizer can make realized savings trail list prices. Above 100,000 tokens Willison says Haiku pricing rises fivefold to $0.50/$2.50, whereas GPT-6 Luna stays at $0.10/$0.50 until 272,000 tokens and then moves to $0.20/$0.75. Reported scores include OSWorld 2.1 offline at 72.4% versus 15.7% for Haiku 4.5, Terminal-Bench 4.0 at 39.2% versus 0.0% previously and 70.6% for Sonnet 5.5, Humanity's Last Exam at 45.9% without tools, FrontierCode 1.1 at 46.4%, GDPval-AA v2.1 at 1,620 versus 735, and a HubSpot CRM-suite score of 92.8%; Latent Space, citing Artificial Analysis, instead reports Terminal-Bench 4.0 at 33% and an Intelligence Index of 43 versus 38 for GPT-6 Luna. Sonnet 5.5 cache reads were halved to $0.10 per million tokens, monthly API credits are offered, cyber safeguards still block penetration testing with biology limits matching larger Claudes, Asana reported over 30% lower latency and up to 2.5x faster turns, sources differ on calling the control adjustable reasoning or the first Haiku with adjustable effort, and Latent Space says Anthropic is fixing a pre-release over-refusal bug. Help Net Security adds that with safeguards off the model achieved code execution in 4 of 410 Chrome V8 runs and finished 3.3% of multi-stage cyber challenges versus 46.1% for Sonnet 5.5 and 67.6% for Opus 5.5, with a 99.71% harmless response rate, Claude Code refusals of malicious requests rising to 84.3% from 66.6%, and stronger prompt-injection resistance than Haiku 4.5 that matched frontier models in most tests but trailed on Gray Swan computer use.

• Anthropic released Claude Haiku 5.5 on October 7, 2026, as claude-haiku-5-5 on the Claude Platform, AWS, Google Cloud, and Microsoft Azure; MarkTechPost and Latent Space cite a 1 million-token context, and Latent Space says Haiku 4.5 had…
• Up to 100,000 tokens: $0.10 input, $0.50 output, and $0.01 cache reads per million tokens; Anthropic says about 75% cheaper to run than Haiku 4.5, while The Decoder and Help Net Security say list prices are up to 90% lower than Haiku 4.5's…
• Above 100,000 tokens, Simon Willison says Haiku pricing rises fivefold to $0.50/$2.50, while GPT-6 Luna stays at $0.10/$0.50 until 272,000 tokens and then moves to $0.20/$0.75; a less generous tokenizer may cut realized savings.
• Scores include OSWorld 2.1 offline 72.4% versus 15.7% for Haiku 4.5, Terminal-Bench 4.0 39.2% versus 0.0% previously (Latent Space/Artificial Analysis says 33%), Humanity's Last Exam 45.9% without tools, FrontierCode 1.1 46.4%, GDPval-AA…
• Sonnet 5.5 cache reads were halved to $0.10 per million tokens; monthly API credits are offered; Asana reported over 30% lower latency and up to 2.5x faster turns; cyber safeguards still block penetration testing.
• With safeguards off, Help Net Security reports code execution in 4 of 410 Chrome V8 runs and completion of 3.3% of multi-stage cyber challenges, versus 46.1% for Sonnet 5.5 and 67.6% for Opus 5.5.
• Help Net Security reports a 99.71% harmless response rate, Claude Code refusals of malicious requests at 84.3% up from 66.6%, and stronger prompt-injection resistance than Haiku 4.5 that trailed on Gray Swan computer use.
• Sources describe adjustable reasoning or the first Haiku with adjustable effort; Latent Space says Anthropic is fixing a pre-release over-refusal bug.