Home Services Work About Blog FAQ Contact
← Back to blog

Claude Haiku 5.5: cheap enough to change what you automate.

Anthropic cut Claude Haiku 5.5's price to $0.10 per million input tokens on October 7 — a steep drop from Haiku 4.5's $1. That's the headline. The number that actually matters for an automation running all day is messier: a new tokenizer eats part of the discount, and the benchmark gap between Haiku and Sonnet is still wide enough that not every job belongs on the cheap tier.

If you've been waiting for Anthropic to make the small model worth running at real scale, October 7 was the day. Claude Haiku 5.5 landed at $0.10 per million input tokens and $0.50 per million output tokens, for any prompt under 100,000 tokens. Haiku 4.5 charged $1 and $5 for the same thing. Push past that 100K mark and the rate steps up to $0.50 input, $2.50 output — still a real cut. Cache reads run a penny per million tokens, cache writes $0.125. On paper, that's close to a 90% price drop on the model most businesses already run their background automations on.

The footnote nobody's pricing table shows

Anthropic buried one detail: Haiku 5.5 runs on an updated tokenizer, and the same piece of text turns into roughly 30% more tokens than it did on Haiku 4.5. That doesn't erase the savings. It just means the sticker price and your actual bill aren't the same number.

A 90% rate cut on a model that needs 30% more tokens to say the same thing isn't a 90% bill cut. Run the real math and it lands closer to 87 — still worth doing, just not the number in the headline.

Which jobs actually belong on the cheap tier

Anthropic built Haiku 5.5 for the boring half of an automation stack: summarizing a thread before it hits a human, compacting context so a longer-running agent doesn't choke on its own history, classifying a support ticket, scoring a lead. None of that needs a model that plans ten steps ahead. It needs one that's fast, cheap, and right often enough.

The early customer numbers back that up, and only for tasks shaped like that. HubSpot ran its own evaluation suite and said Haiku 5.5 "got the best score we've seen on this suite yet, at 92.8% averaged over three runs." AlphaSense tested it across 400 queries and measured a statistically significant jump over Haiku 4.5 — 0.84 against 0.76. Box reported Haiku 5.5 scoring 11 points higher than Haiku 4.5 at about half the latency. Asana measured a 30% drop in latency on task completions and says some agent turns now run up to 2.5x faster.

Ticket triage, lead scoring, invoice categorization, the general cleanup work we described as the "data janitor" agent in an earlier post — that's exactly the zone this fits. Fast, structured, high-volume, forgiving of an occasional miss because a human still checks the edge cases.

Where it still falls over

Push Haiku 5.5 into a job that needs planning across many steps and the gap reopens fast. On Terminal-Bench 4.0, the agentic coding benchmark, Haiku 5.5 scores 39.2%. Sonnet 5.5 scores 70.6% on the same test, and Haiku 4.5 scored a flat zero. On Humanity's Last Exam without tools: 45.9% against Sonnet 5.5's 56.9%. On OSWorld, a test of actually operating a computer: 72.4% against 83.9%. Haiku closed real ground against its predecessor. It still isn't the model you hand a multi-step coding task, or an agent that needs to notice and recover from its own mistakes.

Anthropic's own suggestion matches that gap: run Haiku 5.5 as a subagent underneath Sonnet 5.5 or Opus 5.5 rather than asking it to carry the whole job alone. We made a related point about Sonnet 5.5's pricing a few weeks back — the tiers aren't really competing with each other. They're built to sit in the same pipeline, with the cheap one doing the high-volume filtering and the expensive one handling the parts that need judgment.

Running the actual math

Take a lead-scoring agent processing 500 leads a day: roughly 1,500 input tokens and 200 output tokens per lead, reading the form, checking it against an ideal-customer profile, writing a one-line verdict. On Haiku 4.5's old rate, that's 750,000 input tokens and 100,000 output tokens a day — $0.75 plus $0.50, so $1.25 a day, about $37.50 a month.

Switch to Haiku 5.5 and factor in the 30% tokenizer bump: input climbs to roughly 975,000 tokens, output to about 130,000. At $0.10 and $0.50 per million, that's $0.0975 plus $0.065 — $0.1625 a day, around $4.88 a month. Thirty-seven fifty down to under five dollars is an 87% cut, not the 90% the sticker price implies, but still real money on a workflow that was barely a rounding error to begin with. Run the same ratio on a bigger job — a triage agent handling 5,000 tickets a day instead of 500 — and that 87% is the difference between a few hundred dollars a month and real budget.

What to check before you switch

Swapping in Haiku 5.5 is a model-ID change — claude-haiku-5-5 is live on the Claude Platform, AWS, Google Cloud, and Azure. The one setting worth testing before it goes into production is the effort dial. Haiku 5.5 is the first Haiku-class model that lets you pick Low, Medium, High, Xhigh, or Max effort on the same model instead of jumping to a pricier one. Anthropic's own framing is blunt: it lets you decide whether to optimize for cost or intelligence. The right setting for a lead-scoring agent is almost never the right setting for anything customer-facing.

If you're running Haiku 4.5 on a live workflow, don't just flip the model ID and trust the advertised number. Run your real prompts through both and measure the actual token count. The 30% tokenizer bump doesn't land the same on every kind of text, and that's the only way to know your number instead of Anthropic's average one.

We're already re-pricing a couple of client automations against Haiku 5.5 this month. If you've got a triage or classification workflow sitting on an older model and want someone to run the before-and-after instead of guessing, that's what a scoping call is for.

— Cole

Sources

Running an older Claude model on a high-volume workflow?

30-minute call. We'll benchmark your actual prompts against Haiku 5.5 and tell you the real savings, not the sticker number.

Book a Discovery Call →