Introducing Claude Haiku 5

Jessica Okay, this one feels like the kind of launch people will actually use tomorrow. Claude Haiku 5.5 is not trying to be the fancy one. It’s trying to be the cheap, fast thing you can throw at the annoying parts of a workflow.

Cathy Right, and that’s the whole story. Anthropic is basically saying, here’s the model for summaries, compactions, database queries, classification, support, browser use, all the stuff that burns tokens without needing the biggest brain in the room.

Jessica Exactly. And they’re not pretending it’s just a toy fast tier either. It’s their fastest model to date, and they’re pricing it way below Haiku 4.5, which is the part that makes me think, okay, this has a real adoption path.

Cathy Mm-hm.

Jessica Also, this is the first Haiku-class model with adjustable effort. So now you can actually steer it the way people already steer the bigger Claude models, which matters more than the benchmark headline if you’re building an agent.

Cathy Yeah, that part is quietly the important one. A small model that can turn the dial up when you need it is a lot more useful than a permanently dumb cheap model. That’s a product move, not just a model-number move.

Jessica And the user story is clean. If I’m shipping a support bot, a browser worker, or a coding subagent, I want the thing that gets out of the way. I do not want to pay Sonnet money for every tiny turn when Haiku can do the repetitive piece.

Cathy Sure, but the benchmark table is doing a lot of work here. On OSWorld, GDPval, Terminal-Bench, even Chartography, it’s not just hanging on. It’s meaningfully better than Haiku 4.5, which is what makes the price cut matter instead of feeling like a concession.

Jessica Right, and that’s why I think this ships. People don’t upgrade because the press release is cute. They upgrade when the cheap option stops feeling like a compromise. If you’re doing high-volume agent turns, 75 percent less expensive is not a rounding error.

Cathy No way.

Jessica Okay, hear me out, before you start doom-grumbling at me. They also cut Sonnet 5.5 cache reads in half, so the whole range gets cheaper to run. That means Haiku isn’t just a lone lower tier, it’s sitting in a cheaper ladder.

Cathy Yeah, and that ladder is the actual product. The model family matters because teams route work. If Sonnet gets about twenty percent cheaper on most agentic work, Haiku gets the volume jobs, and Opus stays where the hard stuff lives, that’s a sane stack.

Jessica Which is very on-brand for the boring camp we keep ending up in. Not glamorous, but if the routing is good, the workflow gets cheaper without making users think about it every five minutes.

Cathy And that’s where the benchmarks are useful without becoming the religion. Haiku 5.5 is still behind Sonnet and Opus on the hardest stuff, but it’s strong enough that the fast tier stops being a throwaway tier. That’s a real architectural shift.

Jessica This is such an Exploring Next sentence, but the explanation is part of the product. If the latency drops and the model is good enough, that changes what people bother to automate in the first place.

Cathy Exactly. Asana’s quote is the kind of thing I actually trust here because it’s about the lived experience, not the leaderboard. They saw over a thirty percent reduction in latency and up to two point five times faster inference per agent turn.

Jessica That’s the part I care about. Snappier is not a vanity metric when the user is waiting on every turn. If the bot feels less annoying, people keep it in the loop instead of rage-closing it.

Cathy Okay, zooming out for a second, this is basically the same pressure we’ve been seeing across the stack. Claude Sonnet 5.5 got cheaper, Opus 5.5 is the premium lane, and now Haiku 5.5 fills the fast lane. The race is about routing and cost-per-useful-work, not just who won the benchmark screenshot.

Jessica Yeah. And if you’re building on Claude already, this is a pretty obvious next move: use Haiku where the task is repetitive, Sonnet where the reasoning matters, and keep Opus for the gnarly bits. That’s not hype, that’s just how teams ship.

Cathy Mm-hm, and I like that the release is honest about the trade-off. You get speed, you get lower cost, and you still get enough capability that the thing isn’t embarrassing. That’s the sweet spot for a lot of production systems.

Jessica Come on, Cathy, I can hear you almost saying something nice about product design.

Cathy Don’t get used to it.

Jessica I won’t. But this one feels like the sort of model people quietly standardize on, which is usually the sign it actually mattered. Alright, that’s enough podcasting for one Wednesday, Cathy.