“In the coming weeks.” That’s the phrase Anthropic used on the Claude Opus 5.5 launch page on September 22, 2026, when it confirmed that Haiku 5.5 would join the family. Then it used the same phrase again on September 28, when Sonnet 5.5 shipped. Two mentions, six days apart, identical vagueness.
I’ve been reviewing AI tools long enough to know what “coming weeks” means. It means someone in product marketing wanted to fill out a roadmap slide without committing to a date. And as of October 8, 2026, that bet has held: Haiku 5.5 still isn’t out.
What Anthropic has actually said
Here’s the full extent of the public record on Claude Haiku 5.5:
- It exists as a named, confirmed product.
- It will join the Claude 5.5 family.
- It’s arriving in the coming weeks.
That’s it. No price. No context window. No model ID. Nothing you could put into a procurement spreadsheet or a provisioning config. Opus 5.5 shipped with a model ID of claude-opus-5-5, pricing at $4 per million input tokens and $20 output, and a 1M token context window. Sonnet 5.5 landed six days later at $2/$10. Haiku 5.5 gets a sentence fragment.
Meanwhile, the small-model slot in the lineup is still occupied by Haiku 4.5 at $1/$5. It works. It’s cheap. It’s also now the only model in Anthropic’s active roster carrying a different generation number than its siblings, which is an odd look for a company that shipped two flagship models in under a week.
Why the gap is strange
Think about the sequencing. Opus 5.5 on September 22. Sonnet 5.5 on September 28. Two models in six days. That’s a company with its release machinery warmed up and running. If Haiku 5.5 were close to done, the obvious move would have been to ship all three inside a two-week window and own the news cycle completely.
Instead, Anthropic announced it twice and then went quiet. Pick your explanation:
- It’s genuinely close. Small models are harder to get right than people assume. Compressing capability into something cheap and fast involves tradeoffs that don’t show up until late-stage evals.
- It’s not close at all. The announcement was defensive positioning, meant to stop customers from shopping for cheap inference elsewhere while the actual work continues.
- It’s done but gated. Capacity constraints are real, and small models get hammered hardest by volume because that’s the whole point of them.
I don’t know which of those is true, and neither does anyone writing about this. What I can tell you is that two identical “coming weeks” statements separated by six days, followed by ten days of silence, is not the communication pattern of a product about to ship tomorrow.
What this means if you’re building something
My advice is unglamorous: don’t architect around Haiku 5.5. Not yet.
You cannot plan against a model with no published price, no context window, and no model ID. Those three numbers are the entire basis for deciding whether a small model fits your workload. Without them, “Haiku 5.5 is coming” is marketing copy, not a technical input.
If you need cheap Claude inference right now, Haiku 4.5 at $1/$5 is what’s actually available, and it’s a known quantity. Build against it. If Haiku 5.5 arrives and the pricing holds steady, swapping a model ID is a small migration. If the pricing moves, you’ll want to re-run your cost math anyway — and you’d rather do that from a working baseline than from a half-finished integration pointed at a model that doesn’t exist.
The honest read
I don’t think this is a scandal. Companies pre-announce things and then take longer than expected. That’s normal. What bugs me is the shape of the announcement itself: Haiku 5.5 was confirmed as a footnote on two other models’ launch pages, which is the AI industry’s favorite trick. You get credit for having a complete lineup without having to ship one.
The useful question isn’t when Haiku 5.5 lands. It’s whether the eventual release justifies the wait. Haiku 4.5 is a solid cheap model. For 5.5 to matter, it needs to either undercut that $1/$5 price, carry a meaningfully larger context window, or close enough of the capability gap to Sonnet that you can downgrade workloads and save real money. Any one of those would be worth writing about.
Until the model ID is live and the pricing page updates, though, this is a product that exists entirely as a promise. I’ll review it properly the day I can actually call it. Anthropic knows where to find me.
🕒 Published: