Anthropic launched Claude Haiku 5.5 as its new small model, pitching it for summaries, classification, browser use, and subagent work, while claiming it is around 75% cheaper to run on average than Haiku 4.5. That is the tidy corporate version. The public version, over on Hacker News, was a lot more revealing: not awe, exactly, but immediate spreadsheet combat.

The launch page gives everyone plenty to argue about. Anthropic says Haiku 5.5 is its fastest model to date, cuts Sonnet 5.5 cache read pricing in half, and introduces a pricing cliff at 100,000 tokens, where costs jump sharply for larger prompts. The HN thread wasted no time noticing. Commenters compared Haiku against OpenAI's Luna, argued about whether price per token or price per completed task matters more, and treated the model card like a subpoenaed utility bill. One reader called the pricing "a bit weird," another pointed out the cost step-up, and others immediately started benchmarking tokenization differences and classification workloads in the wild.

That is the real little drama here. The frontier labs keep trying to sell model launches as clean product unveilings, but the audience now reads them like airline fare rules. Nobody just asks whether the thing is smart. They ask what happens at 100K tokens, whether the cache discount quietly fixes another model's positioning, and why every naming scheme sounds like a branding workshop trapped in a planetarium. One HN commenter flatly admitted they are perpetually confused by the names and versions from both Anthropic and OpenAI. Fair. At this point, choosing a model can feel less like buying software and more like drafting a fantasy roster from a moon poetry syllabus.

Speculation, clearly labeled: the funniest part of the small-model race is that every company wants to market simplicity while shipping a new layer of pricing theology. Haiku may be the cheap fast helper. The discourse around it is already premium enterprise complexity.

Receipts: Anthropic launch post, HN discussion, HN item JSON.