Claude Haiku 4.5 powers conversational print ordering for global printing provider.

Executive summary

In a 670-run evaluation per model, Claude Haiku 4.5 matched the top overall accuracy (98.51%) at 2.01 seconds per request, against 6.31 seconds for GPT-5.1. It gave up under two percentage points on strict best-match accuracy for three times the speed, and it now runs the conversation and all four supporting agents in production.

Why print specification is expert work

We at Vstorm built Compass for Rendreal, a commercial printing and print e-commerce company. Compass is a conversational agent that runs on Claude Haiku 4.5, Anthropic's small model, called through the Anthropic API. A customer describes a print job in plain words, and Compass returns a priced, fully specified product in the basket. It then handles changes to orders already placed.

This is conversational print ordering without a designer or a sales consultant in the loop. We ran the LLM model evaluation behind the choice of Haiku, engineered the agent and its supporting layers with Pydantic AI, and took the system into production.

Specifying a print job is expert work. Stock and weight, binding, lamination, format and quantity all interact, and each of them moves the price. A choice made in the first minute can double the cost of a run or rule out a finish later. Rendreal's catalog spans more than a billion valid product combinations.

Before Compass, the process ran on people. Customers who were not designers guessed at a specification, abandoned the basket, or asked someone to build the order for them. When a sales consultant stepped in, the consultant read the specification back to the customer before the price was final. Most new users need that help: 70% struggle with choices such as gutter, bleed, paperweight and coating, which a book author is not expected to know.

Rendreal wanted the process open to anyone who can describe what they want to print. The result also had to be a sensible price, reached without a consultant checking the specification by hand, and an order specification that passes validation when the order is taken in.

Why we chose Claude Haiku 4.5

We tested three models on the product selector, the component that turns a free-text description into a catalog product. The test set held 67 cases in three groups: direct product requests, implicit descriptions, and requests for products Rendreal does not sell. Each case ran 10 times against each model, for 670 runs per model.

Claude Haiku 4.5 matched the highest overall accuracy in the group at 98.51% and reached it in 2.01 seconds per request. GPT-5.1 tied on accuracy but needed 6.31 seconds. GPT-4.1-mini, the lower-cost option, was both slower at 2.36 seconds and less accurate at 97.01%. On overall accuracy, every test case returned the same result across its 10 runs for all three models.

Strict best-match accuracy counts a case only when the expected product is the single top pick, not one of the alternatives offered. Here Haiku scored 92.54% against 94.48% for GPT-5.1: under two percentage points given up for three times the speed.

That trade decides the choice in a chat window, where a customer waits for every reply. Anthropic's launch note points to this kind of work, listing "real-time, low-latency tasks like chat assistants, customer service agents" as a fit for the model (Anthropic). Haiku gave the best quality-to-speed ratio of the three, so it became the model for the conversation and for every component around it.

How Compass works

Compass has a flat architecture. One agent on Claude Haiku 4.5 takes the conversation, and every capability it has hangs off it as a tool, which is a function the model can call to read or change data. The tools fall into four groups:

• Product search, which matches the customer's description to the catalog
• Nine component builders, which assemble the specification one part at a time
• Pricing and the basket, which turn the specification into a priced line item
• Order and publication tools, which cover changes after an order is placed

Haiku decides what each request needs, and a single request routinely runs to six tool calls or more. Four supporting agents run on Haiku around the main one: the product selector, the suggestions generator, the confirmation generator and the conversation title agent.

The specification logic draws on a knowledge base that combines print best practices with the choices earlier customers made and were satisfied with. We orchestrated every agent with Pydantic AI, where we work as an official Pydantic implementation partner. Pydantic AI validates tool calls and outputs against typed schemas and retries when a model returns something malformed. For Compass, that check matters at the last step, because the finished specification has to pass Rendreal's order validation.

When a workflow is clearly bounded, we find that one agent with tools is faster to build, easier to audit and quicker to trust than a distributed design (single-agent development).

What Compass changed for Rendreal

Rendreal's customers now reach a priced specification by describing the job, not by knowing the catalog. The same conversational interface handles changes to orders already placed, so a later change to a finish or a quantity follows the same path as the first order.

Claude Haiku 4.5 is the model in the production configuration today, for the conversation and for the four supporting agents. The 67-case evaluation stays with the system, so the same test can run again when new models ship. Because Pydantic AI is model-agnostic, a change of model does not mean rewriting the agent logic.

The same pattern, one conversational agent with tools, evaluated on real requests before launch, applies wherever customers configure complex products. See how we apply it across print and publishing and how our TriStorm methodology takes an agent from use case to production.