Claude Haiku 5.5 for High-Volume Marketing Routing

For teams exploring claude haiku 5.5 marketing workflows, Anthropic’s model is designed for high-volume, latency-sensitive tasks such as classifying, extracting and routing information. It can be worth testing on short, repeatable decisions such as sorting leads or support messages. Message Threads may help with longer, multi-step conversations, but they do not reduce the tokens Claude reads or the API’s token charges.

The practical starting point is to separate two questions: is Haiku 5.5 accurate enough for your routing task, and does your workflow actually involve a conversation that grows over time? For a single message and one decision, test the model’s cost, speed and accuracy. For an ongoing agent session, also test whether Message Threads reduce request overhead and time to first token.

This guide explains the terms, shows how to estimate a sample bill and walks through a safe evaluation plan. If the work feeds into your broader marketing programme, SearchFIT’s resources can help you explore AI search and content operations, while its pricing page outlines the platform plans.

What marketing routing means

Marketing routing is the process of reading an incoming item, deciding what kind of item it is, and sending it to the right person, system or next step. The item might be a lead form, an email, a support message, a customer question or a content request. The decision might be as simple as “send this demo request to sales” or “tag this question as a product comparison topic”.

A routing system often has four parts:

1. Input: the message and any useful context, such as the form fields or customer account details.
2. Classification: a label describing what the message is about.
3. Decision rules: instructions for what should happen for each label, including when to ask a person.
4. Action: the hand-off to a CRM, support queue, content workflow or human reviewer.

A model is the software that interprets the input and produces an answer. A token is a piece of text that an AI model processes. Tokens are not the same as words: a word may be split into multiple tokens, and punctuation or other text can also use tokens. An API is a way for one software system to send a request to another and receive a response. In this case, your application sends a prompt to Anthropic’s API and receives a model response.

A routing model should not be asked to make an open-ended strategy decision when all the system needs is a label and an action. Keep the task bounded. For example, a lead intake workflow could ask whether a message is about SEO, paid media, analytics or something else, then return the selected category and whether a person should review it. The CRM or workflow tool can handle the actual assignment.

What Claude Haiku 5.5 is, and what it is for

Anthropic released Claude Haiku 5.5 on October 7, 2026. Anthropic describes it as its fastest model and positions it for high-volume, latency-sensitive work, including classification, extraction and routing. Its API model ID is claude-haiku-5-5. It accepts text and image input, supports adaptive thinking with an effort setting, and has a 1-million-token context window and up to 128,000 output tokens, according to the Claude Haiku 5.5 model documentation.

Latency means the time a system takes to respond. It matters in routing because a slow decision can delay a sales hand-off, support response or content workflow. “Fastest” is Anthropic’s description of its own model lineup, not a guarantee that every application will respond quickly. Network delays, prompt size, other software and how you configure a workflow also affect the time people experience.

Adaptive thinking is an option that lets the model use a configured level of effort for its response. The model documentation says the default effort is medium. For a simple routing decision, more effort is not automatically better. Your test should establish whether the selected setting produces reliable decisions at an acceptable response time and cost.

A context window is the amount of information a model can process in a request. A large context window can be useful when a task truly needs a lot of source material. It does not mean you should send a huge history for every small decision. Unneeded context can make a prompt harder to manage and, where it increases billed input tokens, increase the cost.

The model’s capabilities make it a candidate for routine routing, not a guarantee that it will suit every marketing workflow. A model can misunderstand a vague inquiry, choose a label that does not exist or miss an important detail. Treat it as one component in a system with rules, testing and human review, not as a replacement for those controls.

Marketing tasks that may suit Haiku 5.5

The most useful starting point is a task with a clear input, a small set of possible answers and a known next step. These are implementation examples based on Anthropic’s stated use cases, not published Haiku-specific marketing case studies.

Lead intake for an agency

An agency might receive a form message such as: “We need help improving organic traffic and reporting on AI search mentions. Can someone explain your options?” A routing prompt could ask the model to classify the message as SEO, AI search, paid media, analytics or other. It could also extract the company name if supplied and mark whether the prospect asked to speak with someone.

The workflow could then assign the message to a relevant team or queue. It should not let the model invent missing details, such as a company’s budget or urgency. If the message fits two categories, or the user asks a complex question, the system can mark it for review rather than forcing a single answer.

B2B SaaS intent sorting

A SaaS company could classify incoming messages as demo request, support issue, renewal risk or general question. A request such as “Our team is reviewing alternatives and wants a walkthrough” may need a sales hand-off. A message reporting a broken feature belongs in support, even if it also mentions a renewal.

That distinction is important because the cost of an incorrect route is not the same for every label. A delayed demo request may be inconvenient. A support problem or a message signalling a cancellation may need faster human attention. Set escalation rules based on the consequence of an error, not just the model’s confidence in its label.

E-commerce support triage

An online retailer could sort messages into order status, returns, product question, complaint or other. The model can classify the issue and direct it to a queue. It should not make a refund promise or take an irreversible action unless the business has designed and tested specific rules for that action.

For example, a customer asking “Where is my order?” may be routed to order support. A customer saying “The item arrived damaged and I need a replacement” may need a returns or complaints queue, with a human checking the details before a remedy is offered. The aim is to help the message reach the right process, not to give the model authority beyond the task.

Content operations

A content team can use routing to sort customer and search questions by topic or funnel stage. A question about what a product does may go to an explainer brief. A question comparing two solutions may go to a comparison page workflow. A question containing a technical problem can be sent to a subject matter expert.

Here, funnel stage means the broad point someone has reached in a buying or information journey. For example, they may be learning about a problem, comparing options or deciding whether to act. These labels can help a team organise work, but they are not proof of what a person intends. Keep a review path for unclear or mixed questions.

How to estimate Claude Haiku 5.5 costs

Anthropic lists prices per million tokens. For prompts up to 100,000 tokens, the listed rates are $0.10 per million input tokens and $0.50 per million output tokens. For prompts over 100,000 tokens, the rates are $0.50 per million input tokens and $2.50 per million output tokens. Anthropic also lists separate prices for cache writes and cache reads in its model pricing documentation.

Input tokens are the text the model receives, including instructions and provided context. Output tokens are the text it returns. To estimate a basic request, multiply the input token count by the input rate and the output token count by the output rate, then add the two amounts.

Here is a worked example using Anthropic’s listed standard rates for a prompt under 100,000 tokens. Suppose a routing request uses 2,000 input tokens and returns 100 output tokens:

• Input: 2,000 divided by 1,000,000, multiplied by $0.10 = $0.00020.
• Output: 100 divided by 1,000,000, multiplied by $0.50 = $0.00005.
• Estimated total: $0.00025 per call.

At 1 million calls with those same token counts, the modelled total is about $250. This calculation applies the listed input and output rates to the assumed token counts. It is not a quote for a particular deployment. Actual use can differ because prompts vary, output length changes, caching may apply, and other workflow costs are not included.

Anthropic says Batch API requests receive a 50% discount on input and output tokens. Batch processing means sending work to be handled as a group rather than expecting each request to complete as an immediate interaction. It may suit backlogs of content questions or records that do not need an instant response. It may not suit a live lead form where someone expects a prompt hand-off. Check Anthropic’s current documentation and your workload before assuming a batch workflow is appropriate.

There is also a migration detail that can change a cost comparison. Anthropic says Haiku 5.5’s updated tokenizer produces approximately 30% more tokens for the same text than Haiku 4.5, with the increase depending on content. A tokenizer is the component that divides text into tokens. Because pricing is per token, comparing the published price per million tokens alone may not reveal the cost of processing the same content across model versions. Measure token usage on representative examples and include task success in the comparison.

What Message Threads do, and what they do not do

Anthropic announced Message Threads in public beta on November 2, 2026. In a standard multi-turn Messages API workflow, an application normally sends earlier conversation history again with each new turn. With a thread, the application sends the conversation once and then sends only the new messages when continuing it. Anthropic says this can reduce time to first token and keep it lower and steadier as a long conversation grows. See the Message Threads documentation for the implementation details.

A multi-turn workflow has more than one exchange. A turn is one step in the conversation, such as a user message followed by a model response. A request payload is the data sent by your software in an API request. Threads can reduce the amount of conversation history your application has to include in later request payloads.

The key distinction is that a smaller request payload does not mean the model reads less conversation. Anthropic says Claude still reads the full conversation on each turn. Message Threads therefore do not reduce token usage, API pricing or rate limits. The feature can reduce the overhead of repeatedly transmitting history and may improve time to first token in long-running sessions, but it is not a token discount. Anthropic explains this distinction in its article on making long-running agents faster with Message Threads.

Time to first token is the time from sending a request until the model starts returning its response. A lower time to first token can make an interactive workflow feel more responsive, even when the full response still takes time to complete. It is different from total completion time and different again from token cost.

Anthropic reports that, for requests between 500,000 and 1 million tokens, Message Threads reduced median time to first token by roughly 35% for most Claude Code users in its example. This is Anthropic’s reported result for very long agent sessions. It is not a forecast for routine marketing classification, a guarantee for your account or evidence of a cost saving. A short, single-message routing call does not have a growing history to resend, so Threads are unlikely to be useful for that pattern.

When a marketing team might benefit from Threads

Consider a workflow where an agent reviews a long customer conversation, checks order information with a tool, reads the result, and then decides whether to ask a follow-up question or hand the case to a person. Each step adds to the history. In this sort of active session, repeatedly including the full history in the application’s request can create overhead. Threads may be worth testing if the session is long enough for that overhead to matter.

By contrast, consider a new lead arriving in a form. Your application sends the form fields and a fixed instruction, receives a category, then sends the result to a CRM. That is a one-shot request: one input and one response. It has no previous conversation to continue. Focus on the prompt, model choice, output length and routing quality rather than adding a thread feature.

Threads are in public beta and require the message-threads-2026-08-12 beta header. Anthropic says they are available through the Claude API, Claude Platform on AWS, Anthropic-managed Claude on Google Cloud and Microsoft Foundry. Anthropic also says threads expire five minutes after the last request. If a continuation cannot be made, an application may need to send the full conversation again. These details are documented in Anthropic’s Message Threads guide and may change as the beta evolves.

There are data-handling limits to consider. Anthropic says Message Threads are not eligible for Zero Data Retention or HIPAA readiness. It says thread contents are encrypted in memory, are not written to durable storage and are deleted no more than one hour after last use. Zero Data Retention, often shortened to ZDR, refers to a data arrangement under which the provider does not retain request data after processing, subject to the applicable terms. If your organisation has specific privacy, security or regulatory requirements, check the current service terms with your internal reviewers before sending customer or prospect information through Threads.

A practical way to build and test a routing workflow

Start with a single task. Avoid trying to classify every kind of marketing communication in one prompt. Define what enters the system, what the model must decide and which action follows. A good task definition makes it possible to judge a result as correct or incorrect.

1. Define a small label set

Write down the categories in plain language. Give each one a short definition and, where useful, an example. Make sure categories do not overlap unnecessarily. If two labels are difficult for a person to distinguish, the model will probably face the same ambiguity.

Include an “other” category and a clear way to mark uncertainty. Do not make the model choose a specific route when the message does not contain enough information. Decide whether uncertain cases go to a general queue, a human reviewer or a follow-up question.

2. Specify a fixed response format

A schema is a defined structure for data, including the fields allowed and the values each field can contain. For example, a routing result could contain category, priority, needs_human_review and reason. Set allowed values for the category and specify whether the review field is true or false.

Only request fields that your next system needs. If a workflow needs a category and a review flag, a long written explanation may add no value. A short, fixed response is also easier to validate than a free-form paragraph. Have your software check that the response follows the expected structure before taking action.

A sample instruction might say: “Choose one category from the allowed list. If the message does not clearly fit, choose other and set needs_human_review to true. Do not infer details that are not in the message. Return only the requested fields.” Then test it on real examples. The wording is a starting point, not a tested prompt or a guarantee of accuracy.

3. Build a representative test set

Collect examples from the actual sources the system will handle: form submissions, support messages or content questions. Include clear cases and difficult ones. A test set is a collection of examples used to check how a system performs. Have people label the examples before comparing them with the model’s output, and resolve disagreements where possible.

Do not rely only on neat, obvious examples. Include short messages with little context, messages that fit more than one category, spelling variations, unusual requests and examples that should go to a human. Keep examples representative of the languages and channels the workflow will serve. Avoid placing personal information in test data unless your organisation has approved that use.

4. Measure the kinds of mistakes that matter

Accuracy is the share of tested decisions that match the agreed correct label. It can be useful, but a single overall accuracy figure can hide important differences between categories. If most examples belong to one common category, a system can look good overall while performing poorly on a less frequent but high-impact category.

Check false positives and false negatives as well. A false positive happens when the system assigns a label or flag that should not apply. A false negative happens when it misses a label or issue that should have been detected. In lead routing, a false positive might send a non-sales request to sales. In support triage, a false negative might fail to flag a message that needs urgent attention.

Review the errors by category and decide which ones are acceptable, which need a rule change and which require a person to review the case. Do not move to live automatic routing just because the model returns a well-formed answer. A valid format does not prove that the answer is correct.

5. Compare cost and speed with task quality

Log the model, prompt version, input and output token counts, response time, retries, routing result and whether a person changed the result. A retry is another request made after a failure or an incomplete response. Retrying can affect latency and usage, so include it in your measurement.

Compare performance on the same examples, using the same label definitions and success criteria. If you are replacing another model, include both the token count and the price for the relevant version. For Haiku 5.5, Anthropic’s reported tokenizer difference from Haiku 4.5 makes that especially important. The best choice is the model that meets the task’s quality and operational needs at a cost your team accepts, not necessarily the one with the lowest listed rate per million tokens.

For multi-turn agents, measure time to first token and overall completion time with and without Threads. Keep the task, history and test conditions as similar as possible. Do not attribute a difference to Threads if the prompts, model settings or other workflow steps changed at the same time. Keep cost as a separate measure because Threads do not change token usage or API pricing.

6. Add human review and a recovery path

Use human review for ambiguous cases, high-impact decisions and anything your evaluation shows the model handles inconsistently. Set the escalation rules before the workflow is live. A confidence score, if you use one, should be tested against actual correctness; do not assume that a confident-sounding response is a reliable measure of certainty.

For a Thread-based workflow, keep your own copy of each message in the application so it can recover if a thread expires or cannot continue. Anthropic advises applications to keep their own message history for this reason. Build and test a fallback that can reconstruct the conversation when needed. That helps protect workflow continuity, but it does not change what the model reads or what tokens are billed.

What published examples can and cannot tell you

Anthropic has published customer examples that illustrate adjacent use cases. On simulated CRM tasks, Anthropic reports that Haiku 5.5 scored 92.8% averaged over three runs. Anthropic also says that, on a task to identify stale but ambiguous records, HubSpot found it was the fastest among the models it tested and had the highest hit rate and lowest false-positive rate. This is a vendor-published report about CRM tasks, not a controlled comparison of marketing routing systems. See Anthropic’s Haiku 5.5 announcement.

Anthropic also reports that AlphaSense’s “Ask in Document” feature handles about 8 million calls a week. On a 400-query test, AlphaSense reported a score of 0.84 for Haiku 5.5 versus 0.76 for Haiku 4.5. That is an example of a production-scale document question-answering workload. It does not establish the accuracy, cost or speed you will get on your own labels and messages.

These examples are useful for understanding the kinds of work teams are evaluating. They should not replace a test on your own data. The content of a customer’s task, its definition of success, its prompt and its evaluation method all affect the result. A score from a document question-answering test does not directly predict lead classification performance.

Does routing help a brand appear in AI search?

Using Haiku 5.5 for internal marketing operations does not, by itself, make a brand more likely to appear in ChatGPT, Perplexity, Claude, Gemini or Google AI Overviews. Routing can help a team organise incoming questions and spot repeated themes. The next step is to use those insights to create or improve genuinely useful pages, make information easy to access and monitor how the brand is represented in AI answers and search.

For example, a content team could group recurring product questions, check which are already answered clearly on its site, and prioritise gaps for an expert to review. The model’s categories are a way to organise this work, not evidence that a particular page will be cited. A human should verify important questions and ensure that resulting content is accurate and useful.

SearchFIT focuses on AI search growth and marketing workflows. If you are comparing approaches to managing this work, see SearchFIT vs agencies or browse the SearchFIT blog for related guidance. Those resources are separate from Anthropic’s API and do not change the model’s performance, pricing or data terms.

The decision in brief

Claude Haiku 5.5 is a candidate for high-volume marketing routing when the work is made up of short, bounded decisions and the model passes a test on your actual examples. Start with a small label set, constrain the response format, measure errors as well as token use, and send uncertain or consequential cases to a person.

Message Threads address a different problem. They can reduce repeated request payload overhead in an ongoing conversation and may improve time to first token for long-running sessions. Anthropic says they do not reduce token usage, pricing or rate limits. They are generally not relevant to a one-shot lead or message classification call.

Before expanding a workflow, check current documentation, pricing and privacy terms, then run a measured pilot. For the marketing work around AI visibility, explore SearchFIT’s resources and pricing as a next step. Keep the operational question clear: choose the model and API pattern that meet the workflow’s quality, speed, cost and data requirements.