Haiku 4.5: The Model Nobody Expected to Care About

When Sonnet 4.5 dropped, I felt... nothing.

I've been running on Sonnet models since 3.5. Over 40 billion tokens at this point, mostly on Sonnet 4. That's my baseline. That's what I know. So when 4.5 came out with its updated knowledge cutoff and the usual "improved reasoning" claims, I gave it a shot and waited for the moment where I'd think: "Okay, this I couldn't do with Sonnet 4."

That moment never came.

My work doesn't need the latest intellectual facts. It needs reliable reasoning, clean code generation, and solid tool use. All things Sonnet 4 already handled well. A couple weeks ago, I talked this through with someone deeply embedded in the Copilot product — arguably one of the most knowledgeable people at Microsoft on agentic coding. His take? The quality gains we're seeing now come far more from better agent prompts than from model upgrades.

That settled it. I had nothing interesting to say about Sonnet 4.5, so I didn't write anything.

Then Haiku 4.5 came out.

The Model Nobody Heard Of

Most people don't know Haiku exists. It's Anthropic's "small" model — the one Claude Code used quietly in the background to generate conversation titles. It wasn't even available on GitHub Copilot's model list. Why would it be? It wasn't relevant for coding agents.

Then Anthropic released Haiku 4.5 and claimed something that sounded too good to be true: Sonnet 4-level capabilities, at one-third the price, and 2-3x faster.

On SWE-bench, it actually beat Sonnet 4 by a bit.

I've learned to be skeptical of these claims. Benchmarks rarely match real-world use. But I tried it anyway.

The Speed You Won't Go Back From

The first thing you notice isn't the quality. It's the speed.

It's fast enough to change the way the tool feels. Fast enough that the latency disappears and you start hitting keyboard shortcuts without thinking. Fast enough to create the kind of binding where going back feels wrong.

When a tool responds instantly, you stop planning around its delays. You just use it.

The Quality Bar

Here's the part that matters: the quality is good enough.

Not "better than Sonnet 4" in some transformative way. But solidly at the Sonnet 4 level — the benchmark my expectations are tuned to after 40B+ tokens. It handles the same complexity, writes the same quality code, uses tools reliably.

I changed my Claude Code config to make Haiku 4.5 the default. I've been using it for everything since.

A week ago, I shared this with a few colleagues — all heavy Sonnet users. I got exactly one complaint: it doesn't know Hebrew well enough. If you're working in Hebrew, you'll notice the gap.

Why This Matters

This isn't about Haiku being "better." It's about the fact that for most real work, a smaller, faster, cheaper model now matches what the flagship did six months ago.

That gap — between what we think we need and what actually gets the job done — is closing faster than most people realize.

Sonnet 4.5 didn't feel like a leap because the work I do doesn't need one. Haiku 4.5 feels like a leap because it showed me I'll need Sonnet far less going forward.

You should try it.