Anthropic confirmed Claude Haiku 5.5 in a single sentence, tucked into the announcement for a completely different model. The September 22, 2026 Opus 5.5 post introduced Opus as “the first model in our new Claude 5.5 family” and noted that Sonnet 5.5 and Haiku 5.5 were coming soon. That sentence is still, as far as I can tell, the entire public record of Haiku 5.5’s existence.
Six days later, Sonnet 5.5 shipped. Haiku didn’t. And as of late September 2026, there’s no date, no price, no model card, and not a single benchmark number attached to it.
What Anthropic has actually said
Stripped of speculation, the confirmed picture is thin:
- Claude Opus 5.5 — released September 22, 2026, $4/$20 per million tokens, 1M token context
- Claude Sonnet 5.5 — released September 28, 2026, $2/$10 per million tokens
- Claude Haiku 5.5 — confirmed to exist, undated, unpriced
- Claude Haiku 4.5 — the current small model, released October 15, 2025, $1/$5 per million tokens
Fable 5.1 sits above all of them at $10/$50. Two models in six days, then a gap. For a company that has been shipping on a fairly tight cadence, the asymmetry stands out.
Why an undated small model is more interesting than it sounds
If you only read model launches for the frontier scores, Haiku is the boring one. If you actually build agents, Haiku is the one you care about most.
The economics make this obvious. Haiku 4.5 costs $1/$5 per million tokens. Sonnet 5.5 costs $2/$10. Opus 5.5 costs $4/$20. Fable 5.1 costs $10/$50. When your agent loop fires forty tool calls to answer one question, the difference between $1 and $10 on input is the difference between a product and a demo. Small models are where agent pricing lives or dies, and most teams running Claude in production are routing the bulk of their calls to the cheapest model that clears their quality bar.
So the model everyone building at scale needs updated is the one with the least information attached to it. That’s a strange place for Anthropic to leave its customers, even temporarily.
What the gap might mean
I’m not going to pretend I know why Haiku 5.5 is late, because Anthropic hasn’t said. What I can do is lay out the plausible readings and let you weigh them.
Small models are genuinely harder to land
Compressing a frontier generation’s behavior into a cheap, fast model is not a scaled-down version of the same job. It’s a different job. Distillation targets shift, latency budgets bite, and the quality floor you can advertise is harder to guarantee when there’s less model to work with. A delayed Haiku may simply mean the team isn’t happy with it yet, which is the outcome you should want.
The announcement got ahead of the roadmap
Naming three models in a family post is a commitment. It’s also marketing. Confirming Haiku 5.5 on day one of the 5.5 cycle made the family look complete before it was, and now the calendar is doing the talking.
Pricing is the open question
Haiku 4.5 anchored the low end at $1/$5. There is no confirmed price for Haiku 5.5. Whether Anthropic holds that line, raises it to reflect a stronger model, or restructures the tier entirely is the detail I’d watch hardest. A Haiku that gets smarter and more expensive solves a different problem than a Haiku that gets smarter and stays cheap.
My honest read
Nobody should be writing buying guides for a model with no benchmarks, no price, and no ship date. Treat anything you read claiming to know Haiku 5.5’s scores as fiction, because the model card doesn’t exist yet.
If you’re building now, build on Haiku 4.5 and assume it’s your floor for a while longer. It’s been in production since October 2025, which means you know its failure modes, and that’s worth something. Keep your model identifiers in config rather than hardcoded, size your cost projections against $1/$5, and don’t architect around a model that currently has nothing but a name.
The more useful signal here isn’t the delay itself. It’s that Anthropic is now shipping the expensive models first and the cheap one last. For a company whose agent story depends on cheap tokens, that ordering says something about where the attention is going. I’d rather Haiku 5.5 arrive late and good than early and apologetic. But I’d also like a date.
🕒 Published: