Which AI Model Hallucinates Least for Customer Support?

I built a simple n8n workflow that takes an incoming customer question, routes it to a chosen LLM, and logs the response. Then I ran the same set of support tickets through three different models to see how they'd actually perform, not just how they're marketed.

The Setup

A straightforward pipeline: webhook receives a support query → routes to the selected model via API → logs the response for review. I used a fixed batch of realistic support tickets covering shipping questions, refund policy, and product specs — the kind of thing a real support inbox sees daily.

What I Was Measuring

Three things per model: cost per 1,000 queries, hallucination rate (did it invent a policy, price, or date that wasn't in the source material), and accuracy (did the answer match the actual documented policy).

The Pattern

The cheapest model by a wide margin was also the least reliable — it answered fast and confidently, but a meaningful share of its answers included invented details: refund windows that didn't exist, shipping timelines that were made up on the spot. That's the dangerous failure mode for customer-facing automation, because a wrong answer delivered confidently is worse than no answer at all.

The more expensive models were slower to set up (stricter rate limits, more prompt tuning needed) but far more conservative about what they'd assert — when the source material didn't cover something, they were more likely to say so instead of guessing.

The Real Lesson

For anything customer-facing, hallucination rate matters more than per-query cost. A cheap model that occasionally invents a refund policy can cost you far more in chargebacks and support cleanup than the savings on API calls. For internal, low-stakes tasks — like triaging or tagging incoming tickets — the cheaper option is genuinely fine, because a wrong tag just gets corrected by a human, it doesn't get sent to a customer as fact.

Model pricing and relative performance shift often — verify current numbers before citing specific costs.

What I'd Do Next

Add a hard constraint to the system prompt for any customer-facing agent: only answer from the provided knowledge base, and explicitly say "I don't know, let me connect you with support" when the answer isn't there. That one change does more to fix hallucination than switching models does.