The reverse of the Opus question is: why not run Haiku, keep more margin, and increase the WFP donation? On the surface it's an appealing pitch — but the answer is that quality would drop noticeably for most workflows, and quality is the reason people subscribe in the first place.
The quality gap
Haiku is meaningfully faster and cheaper (roughly 1/5 the token cost of Sonnet). It's also meaningfully less capable — noticeably worse on multi-step reasoning, on long-context recall, on nuanced writing register, and on complex code. For daily-driver AI subscription usage, the quality gap matters within the first 20 minutes of trying it. Users would notice.
What Haiku is genuinely good at
- High-throughput API workloads where each individual call is short and simple (classification, extraction, transformation).
- Real-time interactive agents where latency is the primary constraint (voice agents, autocomplete-style features).
- Cost-sensitive backend pipelines processing millions of items.
- First-pass filtering before routing to a bigger model.
Why those uses don't fit LADLE
LADLE is a consumer chat product where each message benefits from the model taking its time to reason. Users write one prompt at a time, wait for the response, then think. Latency is already fine; incremental speed doesn't unlock new workflows. And the failure modes of running a smaller model on chat — subtly wrong answers, poor register, weaker reasoning — are exactly the failure modes that would drive cancellations.
Would the extra donation offset it?
If we ran Haiku, we could plausibly increase the WFP donation from $8 to $12 per subscription. But this only helps if subscribers stay subscribed — and quality degradation is the single biggest driver of AI-subscription churn. Losing 20% of subscribers to get 50% more donation per remaining subscriber is a net loss for the meal-donation math. It also breaks trust: subscribers who tried LADLE at Sonnet and stayed for the quality would be paying for a downgraded product.
If you specifically need Haiku
You'd use Anthropic's API directly, not a consumer product. Haiku is optimized for backend integration; running it in a chat product where you type one prompt at a time and read the response wastes most of its advantages.
When we'd revisit
If Anthropic ships a Haiku-generation model that closes the quality gap with Sonnet (a plausible outcome over 2-3 years as smaller-model training improves), we'd re-evaluate. Until then, Sonnet is the right choice at LADLE's price point.