Haiku 5.5 and the Art of the 75% Haircut

Buried in Anthropic’s own platform documentation for Claude Haiku 5.5 is a line about the model’s training cutoff: it should respond like “a highly informed individual in June 2026” would, and for anything that might post-date that, Claude “often can’t know.” That is Anthropic telling you, in its own words, where the edges of this thing are. I appreciate it. It is more candid than most launch material, and it is a useful reminder that a cheap model is still a model with a horizon.

Because cheap is the actual headline here. Anthropic released Haiku 5.5 on October 7, 2026, as the latest entry in the Claude 5.x family, and it priced the model 75% below Haiku 4.5. Not 20%. Not a quiet adjustment buried in a pricing page diff. Three quarters off the previous generation of the same tier.

What actually shipped

The verified specifics are short, so I will not pad them:

• Haiku 5.5 landed October 7, 2026, joining the Claude 5.x lineup.
• It is multimodal, which pushes Anthropic’s small tier past text-only.
• It ships under a proprietary license. No weights, no self-hosting, no surprises there.
• It sits alongside Sonnet and Opus in Anthropic’s tiered structure, where each model targets a different cost and capability band.
• Pricing is 75% below Haiku 4.5.

That is the confirmed set. Anything beyond it, I have not verified, and I am not going to describe benchmark results I have not seen.

The price cut is the story, and it is not a gift

A 75% reduction on a small, fast model is a competitive move, and the timing is not subtle. Haiku 5.5 arrives during a stretch where cost-per-token has become the loudest axis of comparison in the industry. Other labs have been shipping models explicitly positioned around speed and cost rather than peak capability. Anthropic just made its cheapest tier dramatically cheaper and gave it multimodal input at the same time.

Read that as a bid for volume. The small-model tier is where the boring, enormous workloads live: classification, extraction, routing, summarization, the first pass on a document pile, the triage layer in front of an expensive model. Those jobs are priced by the million, not by the request. A 75% cut changes which of them are economically worth automating at all. Work that penciled out at Haiku 4.5 rates becomes trivially affordable. Work that did not pencil out suddenly might.

It also changes architecture decisions. If you are building agents, the cheap tier is your scaffolding, the thing that handles the dozen intermediate steps nobody sees. Multimodal input at that price makes screenshot handling, document scanning, and image triage viable as routine steps rather than as a premium feature you ration.

What I would not assume yet

Here is where I get annoying. A price cut is a number. Capability is a measurement, and I do not have Haiku 5.5’s measurements in hand.

Small models in a tiered family exist because of tradeoffs. That is the entire point of having Sonnet and Opus above them. Nothing in the verified facts tells me how Haiku 5.5 compares to Haiku 4.5 on reasoning, instruction following, long-context behavior, or tool use. Nothing tells me how its multimodal handling holds up on messy real-world inputs, which is the only kind that matters. A cheaper model that fails more often on your specific task is not cheaper. It is a tax you pay in retries and bad outputs.

So test it on your workload. Not on a demo. Pull a few hundred real examples from whatever you are already running on Haiku 4.5 or a competitor, run them side by side, and compare accuracy before you compare invoices.

My read

This is a solid, unglamorous release doing exactly what a small-tier update should do: get cheaper, get more capable on input types, stay out of the way. The multimodal addition matters more than it sounds, because it removes a reason to escalate to a bigger model.

The 75% figure will get the attention, and it deserves some. Price cuts of that size are not routine, and they tend to signal a lab that wants your default traffic, not just your hard problems. Whether Haiku 5.5 earns that default depends entirely on numbers Anthropic has not put in front of me yet.

Until then, treat it as a promising candidate for your cheap tier and a bad candidate for blind migration. Run the comparison. The savings are real only if the output is.

🕒 Published: October 7, 2026