Claude Haiku 5.5 is the most important model Anthropic hasn’t shipped yet, and the silence around it tells you more about how AI agents actually work than any benchmark chart would.
Let me back that up, because on the surface this looks like a non-story. Anthropic has confirmed Haiku 5.5 exists. It said so in the same breath as Sonnet 5.5, in the September 22, 2026 announcement for Claude Opus 5.5. Sonnet 5.5 then landed on September 28, 2026. Haiku 5.5 did not. As of September 23, 2026, it had no date, no price, no model card, and no published benchmarks. The current small model is still Haiku 4.5, which came out on October 15, 2025.
So we have a confirmed model with nothing attached to it. For most people that’s a shrug. For anyone running AI agents, it’s the detail worth watching.
Why the cheap model is the one that matters
If you’re new to this, here’s the quick version of how Anthropic’s lineup is shaped. There are tiers, and they’re priced per million tokens of text in and out. Right now that looks like Fable 5.1 at $10 in and $50 out, Opus 5.5 at $4 and $20, Sonnet 5.5 at $2 and $10, and Haiku 4.5 at $1 and $5.
Haiku is the bottom of that ladder. It’s the small, fast, inexpensive one. And for AI agents specifically, that position is not a consolation prize. It’s the main event.
An agent is not one question and one answer. An agent loops. It reads a page, decides what to do, tries something, checks the result, tries again. A single task might involve dozens of those steps. Every step costs tokens. So the per-token price of your model doesn’t just add up, it multiplies by the number of steps in your loop.
Run the numbers on the current lineup. On output tokens, Haiku 4.5 at $5 is a quarter the cost of Opus 5.5 at $20, and a tenth of Fable 5.1 at $50. A workflow that’s financially painful on the big model is unremarkable on the small one. That’s why so many real agent systems use a cheap model for the grunt work and save the expensive model for the handful of decisions that genuinely need it.
Which means an upgrade to the cheap tier changes the shape of what’s practical. If Haiku 5.5 is meaningfully more capable than Haiku 4.5 at a similar price, tasks that previously had to be escalated to Sonnet or Opus can stay in the cheap lane. Not because anyone got smarter, but because the floor moved up.
What the gap actually suggests
I want to be careful here, because I don’t know why Haiku 5.5 is late and neither does anyone else writing about it. Anthropic confirmed it and said nothing more. Opus went first on September 22. Sonnet followed six days later on September 28. Haiku was named in the same announcement and is still waiting.
What I’d point out is that the order makes a certain kind of sense. The big model establishes what the family can do. Anthropic’s own framing for Opus 5.5 was that it performs at the level of Fable 5.1 on most benchmarks, which is a statement about the ceiling. Compressing that into something small and cheap is a different engineering problem, and arguably a harder one, because the constraint isn’t “be good,” it’s “be good while staying cheap enough to run in a loop a thousand times.”
That’s my read, not a fact. The fact is that Haiku 4.5 is nearly a year old at this point and still the current option at its tier.
What this means if you’re building something
A few practical thoughts for the non-technical readers who come here, which is most of you:
- Don’t wait for it. Haiku 4.5 is shipping, documented, and priced at $1 in and $5 out. An agent you build today on 4.5 is an agent that exists.
- Write your setup so the model is easy to swap. Keep the model name in one place in your config, not scattered through your prompts and code. When 5.5 arrives, changing one line beats rewriting a project.
- Be skeptical of any article that gives you a Haiku 5.5 release date, price, or benchmark score. None of those have been published. If you see them, someone is guessing and not telling you they’re guessing.
- Track your own costs now. Knowing what your agent spends per completed task on 4.5 is the only way you’ll be able to tell whether 5.5 is actually an upgrade for you, rather than for a benchmark.
The flashy releases get the coverage. The cheap one quietly decides which ideas are affordable enough to build. That’s why an undated, unpriced small model is the thing I’m watching.
🕒 Published: