# Claude Haiku 4.5 on the Cotera agent benchmark

> 4/5 with a faceplant on the easiest benchmark — wrote "I'll help you research..." then logged off.

Source: https://cotera.co/benchmarks/models/claude-haiku-4-5

---

Anthropic's Haiku 4.5 — the smallest, fastest, cheapest Claude. Pitched as a workhorse for high-volume agent loops where Sonnet's bill is too high.

**Score:** 4/5 benchmarks passed · **Cost across the matrix:** $0.941

## Strengths
- Where it didn't fail, it was fine. Reddit, CX, Stripe, and Apollo all passed cleanly with reasonable cost.
- Most economical Coding pass in the Anthropic family — Opus and Sonnet both spent more than Haiku to write the same webhook verifier.
- Total spend was $0.94 — competitive with the cheaper 5/5 models on price, if you can route around the one failure mode.

## Watch-outs
- Failed Sales by writing "I'll help you research Hightouch..." as the entire response. No tool calls, no JSON, no answer — just an opening line. Cost $0.007 to produce nothing.
- This is a known small-Claude pattern: when the system prompt is polite-helper-shaped, Haiku occasionally treats it like a customer-service exchange instead of an agent loop. The rubric punished it; production would too.
- Stripe cost $0.632 — anomalously expensive for a Haiku, more than every Coding pass except Sonnet and Opus. The model went thinking-heavy on a task it should have breezed.