Claude Haiku 5.5

The $0.10 input rate applies to prompts of 100,000 tokens or less, which Anthropic says covers around 90% of requests to the previous Haiku model. Cached input drops to $0.01 per million tokens, so a repeated system prompt costs almost nothing after the first call. That combination is what makes Claude API pricing competitive for classification and extraction at scale.

Haiku 5.5 is the first Haiku class model with an adjustable effort control. Developers can trade accuracy for cost on routine requests and raise the setting when a task needs more thought. Anthropic publishes results at each level, so the choice is measurable rather than guesswork.

The window holds about 1.5k pages of text, and the model can write up to 128,000 tokens in a single reply. Long contracts, support transcripts, and code repositories fit without chunking. Prompts above 100,000 tokens cost five times more, so the practical move is to trim input where accuracy allows.

Calling it a fast AI model undersells the gap. This is the quickest thing Anthropic ships, and that speed is why it works for real time support and for browser agents that read pages step by step. Slow models break the rhythm of a live conversation. Haiku 5.5 keeps pace with a human typing.

The model reads screenshots, charts, and scanned pages as well as text. On Chartography it scores 46.4% without tools, well ahead of Haiku 4.5 at 6.4%. For document intake that mixes scans with typed text, one model handles both streams.

Anthropic recommends pairing Haiku 5.5 with Sonnet 5.5 or Opus 5.5 rather than replacing them. This cheap AI model runs search steps, file reads, and compaction while the larger model holds the plan. Token heavy but simple work moves off the expensive model, and the bill follows.