Anthropic launches new version of scaled-down “Haiku” model (techcrunch.com)

Anthropic has released Claude Haiku 4.5, a new, scaled-down model that the company says matches the capability of Sonnet 4 while running at roughly one-third the cost and more than twice the speed. In Anthropic’s internal tests Haiku 4.5 scored 73% on SWE‑Bench Verified and 41% on the command-line-focused Terminal‑Bench—numbers below Sonnet 4.5 but on par with Sonnet 4, GPT‑5 and Gemini 2.5. Anthropic also reports comparable performance to those models on tool use, computer interaction, and visual reasoning benchmarks, positioning Haiku as a high-throughput, lower-cost option rather than a top-tier reasoning model. The model will become the default for Anthropic’s free plans and is targeted at latency-sensitive, cost-constrained deployments—think large fleets of lightweight agents, parallel sub-agents, or free-tier product experiences where server load matters. Anthropic’s CPO frames Haiku as part of an “agent toolbox” where Sonnet handles complex planning and Haiku executes fast sub-tasks, enabling hybrid deployments (e.g., Sonnet orchestrator + Haiku executors). Immediate use cases likely include software-development tools and other real-time assistants (Claude Code already being common), where the speed/cost trade-off unlocks new production patterns. Haiku 4.5 follows recent launches of Sonnet 4.5 and Opus 4.1, continuing Anthropic’s cadence of iterating across model sizes.