Haiku 5.5 Is The Model I’m Already Architecting Around

Remember October 15, 2025? Haiku 4.5 landed, and within about a day half my side projects got quietly rewired. The summarizer tier, the intent classifier, the little webhook that tags support tickets before a human sees them — all of it dropped down a model class and the bills dropped with it. That was the moment Haiku stopped being the “cheap fallback” in my head and became the default workhorse, with bigger models reserved for the stuff that actually needed them.

Which is why the current situation is a bit funny. Anthropic has confirmed Claude Haiku 5.5 exists. That’s it. No date. No pricing. No model card. No benchmarks. The confirmation arrived as part of the same sentence that confirmed Sonnet 5.5 in the September 22, 2026 Opus 5.5 announcement, and Sonnet shipped on September 28. Haiku is still out there somewhere, due soon, unscheduled.

What We Actually Know

The 5.5 family rolled out fast. Opus 5.5 on September 22 at $4/$20 per million tokens with a 1M token context window. Sonnet 5.5 six days later at $2/$10. Two models in under a week, and then a pause. Haiku 4.5 is still the current small model at $1/$5, which means the lineup right now is top-heavy: Fable 5.1 at $10/$50, Opus 5.5, Sonnet 5.5, and a Haiku that’s a generation behind its siblings.

For bot builders, that gap matters more than it might look. The small model is where the volume lives. If you’re running classification, routing, extraction, or short-form generation at any real scale, Haiku is doing the majority of your calls. An unreleased small model is the one hole in your stack you can’t plan around with numbers, because there are no numbers yet.

Why I’m Not Waiting

The temptation when a model is “due soon” is to freeze. Hold off on the refactor, don’t commit to the prompt library, wait and see what the benchmarks say. I’ve done that before and it cost me weeks of shipping time for nothing.

Better plan: treat the unknown model as a design constraint and build for the swap. Concretely, that means a few things in my own projects.

• Model IDs live in config, never in code. If claude-haiku-4-5 appears inline anywhere in your repo, you’ve already created migration work for yourself. One environment variable, one place to change.
• Every tier gets a benchmark use before it gets a model. I keep a few hundred labeled examples per task — intent classification, field extraction, tone checks. When a new model drops, I’m not reading someone else’s benchmark post, I’m running my own eval in twenty minutes.
• Prompts get written for the task, not the model. Small-model prompts tend to accumulate workarounds: extra few-shot examples, rigid output scaffolding, retry logic for malformed JSON. Tag those workarounds in comments so you know what to strip when the model gets better at following instructions on its own.
• Cost assumptions stay soft. Haiku 5.5 is unpriced. Haiku 4.5 sits at $1/$5. Opus and Sonnet both came in cheaper than Fable 5.1 in this generation, which is interesting, but I’m not building a business case on a price that hasn’t been published.

The Routing Question

There’s a second-order effect worth thinking through. Sonnet 5.5 at $2/$10 is twice the price of current Haiku. If you’ve got a task sitting right on the line between tiers, the honest answer today is often “just use Sonnet and stop optimizing.” That calculus changes if a new small model closes some of the capability gap.

So I’ve started keeping a list — a literal text file — of tasks I’ve pushed up to Sonnet purely because Haiku 4.5 couldn’t hold the line. Multi-step tool calls. Anything needing more than shallow reasoning over a long document. When Haiku 5.5 arrives, that list becomes my test plan instead of a vague feeling that I should “try moving some stuff down.”

What I’d Watch For

Three things will tell you most of the story on day one. First, pricing relative to the $1/$5 baseline. Second, the context window, since Fable 5.1, Opus 5.5 and Sonnet 5.5 share a large window and it’s unclear whether the small model follows. Third, tool-use reliability, which is the single biggest factor in whether a model can carry agent work or just classification work.

Until then, the useful move isn’t speculation. It’s making your stack cheap to change. A model you can swap in an afternoon is worth more than a model you guessed right about.

🕒 Published: October 7, 2026