Ten Cents per Million Inputs: Claude Haiku 5.5 Reprices the Customer Journey

Anthropic launched Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, pricing the model about 75% below Haiku 4.5 on average. That makes AI economically plausible inside the small, repeated decisions surrounding customer support, content classification, in-app assistance and live voice interactions—not only the conspicuous tasks assigned to a premium model. The decision for businesses changes from which large processes deserve AI to which moments need frontier intelligence, which can use a small model and when the system must escalate.

This capability is deployed, not a preview

Haiku 5.5 is available now through Anthropic’s platform, Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic describes it as its fastest small model and positions it for summaries, data compaction, database queries, classification, live support and browser use. Reuters separately reported the launch, pricing and intended applications.

The model also introduces an adjustable effort setting to the Haiku class. Developers can trade more reasoning for more cost on a task-by-task basis instead of choosing one fixed operating level for an entire product. That is commercially consequential because a customer-facing system rarely needs the same intelligence for every turn: identifying a language, extracting an order number and resolving an ambiguous complaint are different jobs.

Anthropic reports large benchmark gains over Haiku 4.5, including computer use and knowledge work. Those results are company-reported, not independent guarantees of performance in a specific business. Availability is verified; fit for a particular customer journey still requires testing.

The unit of automation gets smaller

When model calls are expensive or slow, teams reserve them for visible outputs: a finished campaign, a long report or a complex support case. A faster, cheaper model can sit beneath those outputs and handle the less glamorous work that makes them possible.

One customer conversation can contain dozens of machine-sized decisions: detect intent, retrieve a policy, summarize history, flag urgency, translate a phrase, classify sentiment, prepare an escalation and compress context for the next model. Haiku 5.5 does not prove that every decision should be automated. It lowers the economic threshold at which each decision can be considered.

This matters to marketing because customer experience increasingly depends on invisible orchestration. A brand may use a capable model to formulate an important response while smaller subagents prepare evidence, tag the interaction and update the next-best action. The new possibility is not one cheaper chatbot. It is a layered service in which intelligence is distributed across many narrow moments.

Model routing becomes the product

Anthropic itself draws the boundary: Haiku 5.5 suits narrowly scoped, repetitive work, while Sonnet and Opus remain better choices for complex agentic coding and deeper reasoning. The advantage therefore moves from selecting a single “best” model to designing a routing policy.

That policy needs to account for value, ambiguity and downside. Low-risk classification may stay on Haiku. A sensitive complaint, regulated claim or consequential recommendation may move to a larger model or a person. Adjustable effort adds another control, but also another decision businesses must govern.

For OrionPilot, the useful lesson is architectural rather than promotional. Marketing intelligence becomes more practical when every task is matched to the least expensive model that can perform it reliably, with evidence and escalation built around the boundary. The value is not the lowest token price; it is preserving quality while spending premium intelligence only where it changes the outcome.

Cheap errors can scale faster than savings

Low prices invite volume, and volume amplifies weak decisions. A small false-positive rate can misroute thousands of support cases, suppress useful customer signals or personalize the wrong offer at scale. Faster responses can also create a review problem if generation outruns the organization’s ability to inspect what the system is doing.

Pricing has a boundary too. For prompts above 100,000 tokens, Haiku 5.5 rises to $0.50 per million input tokens and $2.50 per million output tokens. It remains cheaper than Haiku 4.5, but the fivefold step means poor context design can erase part of the expected advantage. Businesses need to measure complete task cost, including retrieval, repeated calls, escalation and correction—not quote the cheapest line on the pricing table.

Anthropic cites favorable early tests from Asana, HubSpot, AlphaSense, Box and others. These are useful signals, but they come through Anthropic’s launch materials and do not substitute for independent evaluation across brands, languages, edge cases and live customers.

Watch the cost of a resolved outcome

The evidence businesses should watch is operational: accuracy by task type, tail latency during peak demand, escalation rate, false positives, customer abandonment and cost per correctly resolved interaction. Teams should also track whether cheaper calls reduce total spending or simply create new workloads that were never previously attempted.

Haiku 5.5 is not important because it makes one impressive answer cheaper. It matters because it can make millions of ordinary decisions economically visible. If businesses can route those decisions without multiplying errors, the next competitive advantage will not be owning the smartest model; it will be knowing precisely when not to use it.