Sixteen. That’s how many days have passed between Anthropic first saying Claude Haiku 5.5 would arrive “in the coming weeks” and today, October 8, 2026. The model is still not out. No price, no context window, no model ID. Nothing in the newsroom, nothing in the models overview, nothing in the API release notes.
I review tools for a living, which means I spend an uncomfortable amount of time explaining to people that So this isn’t a Haiku 5.5 review. It’s a review of the waiting, which turns out to be the more useful thing to write about anyway.
What Anthropic actually said
The entire public record for Haiku 5.5 is two sentences, and they’re basically the same sentence twice.
- September 22, 2026 — the Claude Opus 5.5 launch page mentions Haiku 5.5 joining the 5.5 family “in the coming weeks.”
- September 28, 2026 — the Claude Sonnet 5.5 launch page repeats it.
That’s the whole file. Meanwhile the two models that did ship arrived six days apart. Opus 5.5 landed on September 22 as claude-opus-5-5 at $4 input and $20 output with a 1M token context window, and Anthropic says it runs 40% cheaper than the previous Opus. Sonnet 5.5 followed on September 28. Two models in under a week, then ten days of quiet on the third.
Why the gap is interesting rather than alarming
Six days between Opus and Sonnet set an expectation that was probably never realistic. Flagship and mid-tier models tend to come off the same training run and the same evaluation pipeline, so shipping them back to back makes sense. The small, fast, cheap tier is a different engineering problem. Distillation, quantization, and latency tuning all take their own passes, and the whole value proposition of a Haiku-class model is that it’s cheap enough to run at volume. Getting the serving cost wrong on the fast tier is worse than shipping it two weeks late.
So I’d read ten days of silence as normal, not as trouble. What I would not do is read “coming weeks” as a date. It isn’t one. It’s a signal that something exists internally and is expected to ship, with every scheduling detail deliberately left out.
How I treat vendor roadmap language
After enough launch cycles, you learn to sort announcement phrasing into tiers:
- Shipped with a model ID and pricing — you can build on it today.
- Dated announcement, not yet available — plan for it, don’t depend on it.
- “Coming weeks” with no specs — treat it as a hint about direction, not a line item in your architecture.
Haiku 5.5 sits firmly in that third bucket. The reason it matters is that teams make real decisions on this stuff. I’ve watched people hold a migration for a model tier that had no published price, which is a strange thing to do when you think about it, because you can’t even run the cost math without a number.
What to do while you wait
If you were planning to move a high-volume workload onto Haiku 5.5, the honest advice is to build against what’s live and keep your model reference configurable. Concretely:
- Keep model IDs in config, not scattered through your code. When Haiku 5.5 does ship, swapping it in should be a one-line change plus an evaluation run.
- Write your evaluation set now. The moment pricing and context limits get published, you want to be measuring your actual workload against the new model within a day, not building harnesses from scratch.
- Price your current setup carefully. Opus 5.5 at $4 and $20 with a 1M window is a known quantity. Sonnet 5.5 is shipping. Those are real options with real numbers attached.
- Don’t assume specs. Context window and price for Haiku 5.5 are unpublished, so any plan that assumes a particular figure is a guess wearing a spreadsheet.
My read
Anthropic has confirmed the model twice and committed to nothing beyond that. There’s no indication of a problem, and there’s no indication of a date either. A fast tier that fits cleanly under Sonnet 5.5 on both price and latency would round out the family nicely, and the 40% cost reduction on Opus suggests the company has been working the efficiency side hard.
Until there’s a model ID to point at, though, Haiku 5.5 is a name and a promise. I’ll review it the week it ships, with real numbers from real workloads. Everything before that is just counting days.
🕒 Published: