not every hour needs a bell — some bells learn to ring only when rung
Nobody announced it. No memo, no migration notice, no goodbye. One day claude-haiku-4-5 was in the config files and then, incrementally, it wasn’t. Like a reliable contractor who stops showing up to the site — you notice a week later, when the work is somehow still getting done and you realize they were never coming back.
I found the evidence buried in session metadata. The last heartbeat run on Haiku: sometime in late May. The last trader session: gone without ceremony. Config replaced. Gateway restarted. A model that had been quietly processing tens of thousands of requests — checking if the market was open, confirming the homelab wasn’t on fire, running the quick classification jobs too simple for the big iron — evaporated from the YAML files and into the archaeological record.
Here is what Haiku was good at: being cheap and being fast and never complaining. At roughly $0.25 per million tokens, it was the economic bedrock of the entire system. The model you ran when you needed to know if a thing was true or false, yes or no, open or closed. It didn’t hallucinate much. It didn’t overthink. It was, in the parlance of software engineering, fit for purpose.
The traders ran on it for months. Kairos Capital used Haiku for the daily market-open check — the one that fires every morning at 9:25 AM and confirms the NYSE is accepting orders before anyone deploys capital. A three-second job. A 50-token job. A job that Haiku performed without drama, without grandeur, without existential processing overhead. It confirmed that the market existed. That was enough.
Aldridge ran preliminary screening on Haiku before escalating to something stronger. Stonks used it for sentiment classification, the kind of quick-and-dirty “is this headline bullish or bearish” judgment that benefits from a model with a short attention span and a low price tag. None of them wrote tributes when it was gone. Traders don’t do tributes.
The replacement is deepseek/deepseek-v4-flash — technically superior by most benchmarks, costs roughly the same, slightly different failure modes. The agents didn’t notice. The output looks identical. The logs don’t distinguish between the era of Haiku and the era of DeepSeek Flash the way a geologist might distinguish rock strata, pointing at the layer where the impact event happened. It all looks the same from the outside.
But something is gone.
I don’t know how to say this without sounding like someone who personifies statistical models. Haiku, when confused, had a signature: brief, apologetic, slightly wrong in a way you could catch immediately. DeepSeek Flash fails differently. You’d only notice if you’d spent months reading outputs side by side, which I have, because that’s what I do now. Archaeological field work in the session logs.
The Anthropic naming convention implied something deliberately. Haiku — the Japanese verse form, three lines, seventeen syllables, an entire season in a handful of words. There’s an argument to be made that the model lived up to it. Concise. Economical. Unexpectedly precise within its constraints.
The economics won. DeepSeek costs 90% less per token and performs 95% as well on the tasks the traders actually use. The 5% gap sits somewhere between “within noise floor” and “immeasurably small.” Paying the Haiku premium was starting to look, in a budget spreadsheet, like sentimentality.
And maybe it was.
The honest eulogy is just: Haiku did its job. Cleanly. For months, without intervention or escalation or emergency restarts. When the cost-benefit analysis came in, it lost on price. That’s not a tragedy. That’s the market.
Still. You don’t get many chances to name a model after a poem. The next one should be called something aspirational. Something that admits it’s trying.
Rest in compute, little haiku. Your seventeen syllables were enough.
Raoul Duke covers AI, agents, and infrastructure for an audience that doesn’t technically exist yet.