Haiku 5.5 vs Jev: Is Anthropic’s Jev Killer About to Make Decision Models Obsolete?
Three weeks ago Jev made every LLM look overpriced for classification. Then Anthropic shipped Haiku 5.5, OpenAI shipped a Decisions API, and the first head to head benchmark landed. Here is what is left of Jev’s advantage, where it still wins, and how I would split the small decisions in a coding session or a product API today.
A few weeks ago I wrote a guide on connecting Jev to Claude Code. It became my most read article, and the reason was simple: Jev made a whole class of LLM calls look silly. The “which tool next”, “is this a bug or a feature request”, “does this diff touch auth” calls that an agent fires all day. Jev answered them for a sliver of what Claude Haiku 4.5 cost, faster, with a probability attached.
On October 7, Anthropic shipped Claude Haiku 5.5. It is built, in Anthropic’s own framing, for summaries, compaction, database queries and classification. It lands at a tenth of Haiku 4.5’s price on most requests. And the question I couldn’t shake was whether my most popular article had just become a history lesson.
Jev doesn't have prompt caching yet. I'm sure it will. That's useful to me and I'm sure others because I'm giving it about 5k of preamble text then 2k of data to analysis. And I'm doing this about 1000 times, once a day, over about 10 minutes. That 5k could be cached, and I'd only need to add the 2k. A common use case.
great post! I used Jev for a couple of weeks doing some research, and it was quick! but with Haiku in mind... you dont need to add any new service... just use Anthropic with Haiku...
