Claude Haiku 5.5 Cuts Small-Model Costs by 75 Percent

The cheapest model in the room is the one you should watch

Anthropic released Claude Haiku 5.5 on October 7. It is not the kind of launch that trends on its own. Haiku is the small member of the Claude family, the one that does the boring jobs, and small-model releases rarely make headlines. This one deserves a second look anyway, because of a single number: about 75 percent lower running cost than Haiku 4.5.

Three quarters off the cost of an already cheap model changes what is worth doing. When a function is essentially free, you stop rationing it.

Why small models are decisive

Most of the interesting AI systems shipping right now are not one model. They are orchestrations. A large model plans, and several small models handle the individual steps: summarize this document, classify this ticket, answer this database query, check this code change. The big model gets the credit. The small models do the volume.

That means the economics of an agentic product often hinge on the cheap tier, not the flagship. If your subagents cost pennies, you can afford more of them, more retries, and longer chains. If they cost dollars, you design around scarcity, and scarcity shows up as worse products: fewer checks, shorter loops, less willingness to try a second approach.

What Haiku 5.5 actually offers

Haiku 5.5 is priced the same as Haiku 5 at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20. Anthropic says it costs up to 30 percent less for most work and delivers 30 percent more usage inside Claude Code. It also ships with an adjustable effort setting, which lets you dial how hard the model thinks before answering, a useful knob when you are optimizing for cost rather than depth.

Anthropic positions it for summaries, database queries, classification, and coding-subagent work, and notes that its model card shows no crossing of the CB-2 and Autonomy-2 thresholds, with low catastrophic-misalignment risk. For teams that need sign-off before deploying a model, that detail is not decoration. It is the difference between shipping and waiting in a review queue.

The adjustable effort setting is the most underrated part. It lets the same model serve two very different jobs: a fast, cheap pass for high-volume filtering, and a slower, more careful pass for the cases that survive the filter. That pattern, cheap first then careful second, is how most cost-efficient pipelines end up being built.

The 5.5 generation keeps compounding

Haiku 5.5 is the second model in the 5.5 family, following Opus 5.5. The pattern across the industry in October is consistent: reasoning gets a little cheaper, small models get a little better, and the floor under everyday AI work keeps dropping. OpenAI pushed GPT-6 to everyone, Google previewed Gemini 4 Argon and cut prices with it, and Anthropic trimmed the bottom of its own lineup.

None of these moves is dramatic on its own. Together they describe a market where capable intelligence is becoming a commodity input, and the value migrates to everything wrapped around it: the data, the workflow, the evaluation, and the trust. When the model itself gets cheap, the differentiators are the parts that were never the model.

Routing becomes a discipline

When the price gap between a flagship and a small model is large, the highest-value engineering decision is often not which model to use, it is which model to use for which step. A well-built system sends the easy 80 percent of requests to a cheap tier and reserves the expensive model for the cases that actually need it.

That is a routing problem, and it is harder than it looks. You have to judge difficulty before you have the answer, which means building a classifier or a heuristic that decides where each request goes. Get it wrong in one direction and you waste money on trivia. Get it wrong in the other and you degrade quality on the requests that matter.

A 75 percent cheaper small model makes the safe side of that tradeoff wider. You can afford to route more aggressively, and you can afford to catch the mistakes with a verification pass, because both the routing decision and the check are now cheap.

What to do with the savings

The obvious move is to route more work to the small model. The subtler move is to reconsider tasks you had written off as too expensive. A system that reads every support ticket, or labels every image in a catalog, or drafts a first pass on every document, was a budget decision before. At the new price, some of those decisions flip.

There is a trap here too. Cheaper tokens tempt teams to skip evaluation and let low-quality output flow at volume. Cost falling does not make errors free. If anything, it makes measurement more important, because the failure mode shifts from "we could not afford to try" to "we tried everything and nobody checked."

The teams that handle this well will spend part of the savings on exactly the boring infrastructure that cheaper tokens make affordable: sampling, evaluation, and guardrails. That is the difference between using a price cut to scale output and using it to scale quality.

Why the price cut is bigger than it looks

It is worth pausing on the compounding effect of a price cut like this one. Teams do not spend a fixed budget on model calls. They spend against a target for quality, and when each call gets cheaper, they do not pocket the difference. They raise the target. More checks, more retries, better context, longer chains. A 75 percent cut lowers the bill and raises the ambition of what a team is willing to attempt.

That is how earlier price drops played out in practice. When reasoning models got cheaper, systems started reasoning about more steps. When context got cheaper, systems started feeding in more documents. The price of a unit of work sets the shape of the system built on top of it, and a cheaper unit buys more careful design.

The version of this that matters for agents is especially direct. An agent is a loop, and the cost of the loop is the cost per step times the number of steps. Cut the per-step cost hard enough and you can afford to let the loop run longer, which is most of what separates a helpful agent from a frustrating one. Haiku 5.5 is a step toward loops that are worth running to completion.

The quiet argument

Frontier model launches get the attention. But the models that decide what an AI product can afford to do are usually the small ones. Haiku 5.5 is a reminder that a 75 percent price cut on the bottom tier is not a footnote. It is a change in what is economically possible, and those changes tend to show up in shipping products a few months later, in features nobody thinks to credit to a model release.