Remember the stretch in late September when Anthropic shipped Opus 5.5 on the 22nd and Sonnet 5.5 on the 28th? Two models in six days. My Slack was a wall of people swapping model IDs mid-sprint and arguing about whether the pricing math worked out. I rewrote one router config twice that week. It felt like the small, fast model was going to land any minute.
It hasn’t. As of October 8, 2026, Claude Haiku 5.5 is still unreleased. Anthropic has said twice that it will join the Claude 5.5 family “in the coming weeks” — once on the Opus 5.5 launch page on September 22, and again on the Sonnet 5.5 launch page on September 28. Since then: nothing in the newsroom, nothing in the models overview, nothing in the API release notes. No price, no context window, no model ID.
That’s the whole factual picture, and I think it’s more useful to bot builders than another round of speculation.
Why the missing Haiku matters architecturally
If you build agents or chat bots for a living, the cheap-fast tier isn’t a nice-to-have. It’s the layer that handles intent classification, routing, short summarization, tool-call argument formatting, retry passes, and the hundred tiny inference calls that make a bot feel responsive instead of sluggish. Those calls dominate your request volume even when they don’t dominate your reasoning quality.
Right now, for the 5.5 family, that tier is a blank. Opus 5.5 is published at $4 per million input and $20 per million output with a 1M token context. Sonnet 5.5 shipped six days later. The small sibling that would normally absorb the high-volume work is confirmed to exist and otherwise undefined.
So the practical question isn’t “when does it drop.” It’s “how do I build so that the answer doesn’t matter.”
Three things I changed in my own stack
I had hardcoded model strings in four places across one project. That was embarrassing to find and worth fixing regardless of Haiku. Here’s what I did instead.
- Named tiers, not model IDs. My config now exposes
FAST,BALANCED, andDEEP, each mapping to a model ID in one file. When a new model lands, I change one line. The agent code never sees a version number. - Per-tier eval suites. Fifty to eighty recorded real tasks per tier, with pass/fail assertions I actually trust. Not vibes. When a new small model appears, I can answer “is it good enough for routing” in an afternoon rather than a week.
- Cost and latency logging per tier, not per request. If I don’t know what my FAST tier costs today, I can’t tell whether a swap is a win. I was flying blind on this for months.
None of that is clever. It’s the kind of plumbing you skip when you’re shipping fast and regret the moment the model menu changes under you — which, going by the September pace, it will.
What I would not do right now
I would not hold a launch for Haiku 5.5. “In the coming weeks,” repeated twice without a follow-up, is a signal about intent, not scheduling. Anthropic has told us the model is coming. It has told us nothing about when, what it costs, or how much context it holds. Any roadmap you build on those three unknowns is a roadmap built on guesses.
I also wouldn’t pre-optimize for a context window you can’t look up. I’ve watched people design chunking strategies around a number they read in a forum thread. If the spec isn’t published, treat it as unpublished. Build your retrieval layer so chunk size and window budget are parameters, and move on.
And I’d resist the urge to rip out whatever small model is currently doing your high-volume work. It’s working. A model that doesn’t have a model ID yet cannot be benchmarked, priced, or load-tested. The comparison you want to run doesn’t exist as a comparison yet.
The useful version of waiting
There’s a decent habit hiding in moments like this. A gap between “confirmed” and “shipped” is free time to make your system indifferent to which model sits behind each tier. That’s work you’d have to do eventually anyway, and doing it now means the actual launch is a config change and an eval run instead of a refactor.
When Haiku 5.5 does land, the builders who get value out of it fastest won’t be the ones who guessed its price correctly. They’ll be the ones who already had a slot shaped like it. Go build the slot.
🕒 Published: