BLOG
Make’s October 10, 2026 update says Claude Haiku 5.5 is available in its AI modules. Anthropic launched the model on October 7 as an option for high-volume, cost-sensitive work, according to its official announcement. For a small business, the sensible pilot is a bounded task such as classifying a form submission or extracting a few fields for review. Availability in Make does not mean every existing scenario should be switched without testing.
What arrived in Make?
Check whether an older scenario uses a fixed model name or a dynamic selector. A new choice in the catalog may not affect an existing saved configuration. If a teammate changes the model, save the previous selection and run the test set again. That makes the switch reversible and its effects visible.
Make’s announcement describes Haiku 5.5 for classifications, routing, extraction, and other workflow steps. Check the current AI tools documentation for the exact module, model selector, connection route, and fields available in your account. A community announcement can identify the release, but the live module determines what you can configure.
Anthropic’s claims about speed and cost are vendor claims for its model under published conditions. They do not demonstrate the time or cost of your complete Make scenario. A workflow can wait on a CRM, repeat after an error, or require staff review. Measure the entire task.
Choose the module for the task
A simple classification or extraction does not require an autonomous agent. Use the smallest module that returns the fields the next step needs. Define allowed categories and an “uncertain” outcome. For a document, specify which fields must be present and what happens when a scan is unreadable. Avoid telling the model to invent missing values so the scenario can continue.
An agent with tools may be useful for a more variable task, but every connected tool expands what it can do. Review credentials and the data it can read or write. Keep customer-facing messages and CRM updates behind approval until the behavior is well understood. A model label alone does not define the permission boundary.
| Use case | First safe output |
|---|---|
| Support form | Category and review flag |
| Invoice document | Extracted fields with missing-data marker |
| CRM routing | Suggested owner for human check |
| Customer response | Draft, never an unreviewed promise |
Make credits and Anthropic API charges differ
Make’s credits guide covers its platform consumption. An own-key or separate Anthropic API route can have a distinct provider bill. Check the connection selected in your scenario and Anthropic’s current model pricing before estimating a monthly cost. Do not use a Claude application subscription price as the API rate. Include Make runs, retries, other connected services, and staff time.
Set a cost owner and alerts in the accounts that actually pay. A public webhook can receive far more traffic than a pilot. Rate limits, retry configuration, and an outage in the destination CRM can change the usage pattern. Put a maximum queue or manual fallback in place rather than letting a failed flow repeat indefinitely.
Example: route a website form to a CRM
A small services firm receives requests for site creation, maintenance, and automation. The workflow stores the original form and its submission ID. Haiku 5.5 suggests a category and notes missing information. An explicit rule sends ambiguous requests to a person. The CRM lookup checks whether the contact already exists before creating a new lead, using the submission ID to keep retries safe.
Test a normal request for each category, a message mentioning two services, a missing email, a repeat submission, and a CRM failure. Verify the actual saved record and the alert, not only the Make execution log. If the model classifies a maintenance emergency as a general quote, the human review route should catch it before a delayed response harms the customer.
How to compare before replacing a model
Run the new model and the current configuration on the same authorized examples. Measure correct routing, missing fields, review time, latency observed in your scenario, and total cost per accepted result. Do not transfer a vendor benchmark directly to this workflow. Keep prompts and output schemas versioned so you can reproduce a change.
Check what happens when the model returns no category or an invalid field. The next module should reject or route it for review rather than writing a placeholder to the CRM. A small, understandable error path can be more valuable than a slightly faster answer.
Should the first pilot send messages?
Prepare a data contract for each model step
Write a small schema before connecting the model to another module. Name each input field, allowed output category, maximum length, and handling for missing values. For an invoice, a blank tax ID is different from a zero value. For a customer inquiry, “unknown service” is different from an empty string that the CRM might silently replace. Validate the model output in a non-AI step before any record is changed. Keep the original document or form linked to the extracted data so staff can check it.
Use the same contract when comparing Haiku 5.5 with an existing model. Otherwise a new prompt may appear more accurate simply because it receives extra information. Mark which fields the model is allowed to infer and which require a source. In most business records, dates, amounts, and contact details should come from the source or be flagged for review, never completed from a plausible guess.
Watch for drift after a successful pilot
Customer requests, document templates, and model behavior can change. Review a sample of accepted and rejected results every week after launch, then adjust the prompt or routing rule with a version number. A sudden rise in “uncertain” classifications may reflect a new service offering rather than a model defect. Keep a manual route while diagnosing it. Confirm that a changed model selector or connection in Make did not happen silently. The team should be able to say which configuration processed a given CRM record.
Begin with internal suggestions. Staff can correct facts and tone, and the company can learn which requests are ambiguous. Only consider a narrowly defined automatic response after measuring errors and confirming approved templates, privacy rules, and a human escalation route. A new model in Make is an option to evaluate, not a mandate to automate every customer exchange.
