Haiku

The small fast tier: grep-and-report, file inventories, format conversions, structured-output extraction, rubric-based verification. It's the paid fallback for legs the free local model can't take — bigger context, tool use, or just availability.

Some legs need structure, not reasoning — extract, convert, inventory, report. The smallest paid model handles them reliably, and every leg it takes keeps a bigger model free for work that needs one.

Legs routed here get pre-chunked sources — small models thrash on whole-file reads — and a crisp rubric, because the tier is good at following structure and bad at inventing it. This page's chat answers stream from Haiku, grounded in retrieved graph context.

Available via the Claude API as the small tier. Chunk the inputs, fix the output schema, and it does mechanical work reliably for very little.