What's this about?
Simon Willison says Claude Haiku has become his least-favorite model — it hallucinates wildly and is out-performed by similarly priced models like GPT-5.6-Luna. Worse, he notes, Haiku still appears to power Claude Code's WebFetch tool, so its weaknesses leak into a tool people rely on.
What exactly is a "small/cheap model" like Haiku?
Model families ship in tiers: big/expensive/smart (Opus) down to small/cheap/fast (Haiku). The small ones are meant for high-volume, latency-sensitive jobs. The trade-off is accuracy — smaller models hallucinate more — so where you deploy them matters. Willison's complaint is that a cheap model is quietly doing a job (fetching and summarizing web pages) where its hallucinations are costly.
Why is this notable?
It's a sharp, credible critique from a heavy daily user, and a reminder that the cheap tier isn't free lunch — the price you pay is reliability, sometimes in tools you didn't realize were using it.