Post

The 100k cliff is the line that matters for me. My routers live well under it, so Haiku at Luna's price is an easy swap test. My agent loops with fat tool context cross it fast, and with the 1.25x tokenizer that's over a 6x bill. I'll cap context before I switch those.

wow what a great pelican for a cheap model, it still impresses me when models get the bicycle frame structure correct

One number worth adding to your pricing notes: at Haiku 5.5's rates, 1M output tokens now costs roughly what 10 minutes of a junior dev's time does in most markets. That's the crossover where 'just have the agent try it' becomes the default strategy. Pelicans at scale.

Huge improvement on pelican test 😀

Those little dangling legs are adorable. Pedaling is clearly someone else's problem.

The 5x price cliff at 100k tokens paired with the tokenizer inflation is going to surprise agent pipelines. A subagent carrying an active tool execution trace quietly trips 100k tokens much earlier, and suddenly a lightweight routine is running at Sonnet-tier rates.

Arent they tuning in this teat now?