Behind Claude Haiku 5.5's cost reduction: big gains in agentic ability, but complex coding still trails Sonnet 5.5

Anthropic releases Claude Haiku 5.5; its OSWorld 2.1 score rises from 15.7% in the previous generation to 72.4%, and average operating costs drop by about 75%. The new model introduces Adaptive Thinking and Effort strength settings to the Haiku line for the first time, and supports a 1 million token context. However, it scores only 39.2% on Terminal-Bench 4.0, far below Sonnet 5.5's 70.6%; Anthropic still recommends Sonnet 5.5 or Opus 5.5 for complex agent coding tasks.