Claude Haiku 4.5 System Card [pdf] (assets.anthropic.com)

Anthropic published a system card for Claude Haiku 4.5, a new “small, fast” hybrid-reasoning LLM optimized for coding and computer use. Haiku 4.5 is smaller and faster than Anthropic’s recent flagship models, trained on a proprietary mix of public web data (up to Feb 2025), non‑public third‑party sources, opted‑in user data, and internally generated examples, with posttraining fine‑tuning via human and AI feedback. Key model features include a user‑toggleable “extended thinking” mode that exposes a chain‑of‑thought, explicit context‑awareness (helping the model manage a 200K‑token window to avoid premature stopping), and improvements aimed at multi‑instance agentic workflows and agentic coding. The system card emphasizes extensive safety testing across single‑turn and multi‑turn harms, agentic safety, reward‑hacking, bias, child safety, and biological/cyber risk benchmarks. In single‑turn violative request tests Haiku 4.5 produced harmless responses 99.38% (±0.21%) of the time, comparing favorably to Claude Opus 4.1 and Sonnet 4.5 and showing marked gains over Haiku 3.5. Anthropic’s Responsible Scaling evaluations ruled out ASL‑3 thresholds, and Haiku 4.5 was deployed under the ASL‑2 standard. For practitioners, this signals that a more capable, low-latency model family is now available with substantial safety engineering and monitoring, but Anthropic notes some residual edge‑case exceptions and continues ongoing adversarial and welfare testing.