Back
Reports & Research

Inference Cost Benchmarks for Customer Support Workloads

Research · PDF10 min readUpdated Aug 2026Neura Dynamics

We ran the same support-ticket workload across open and proprietary models and compared answer quality, latency and cost per resolved ticket.

Method

Each model answered the same anonymised set of support tickets using an identical retrieval setup. Answers were scored by trained reviewers against a shared rubric.

How to use the results

The cheapest model per token is rarely the cheapest per resolved ticket. The report shows where smaller models are good enough, and where escalation to a larger model pays for itself.

01

Methodology and dataset

We used an anonymised set of support tickets with known resolutions. Every model received the same retrieved context and system prompt.

02

Quality scores by model

Larger models scored highest on complex multi-step tickets. On routine tickets the gap between large and small models was small.

03

Latency at p50 and p95

Median latency varied widely between providers. Tail latency at p95 mattered more for user experience than the median.

04

Cost per resolved ticket

Cost per resolved ticket combines token cost with the rate of escalation to a human. A cheaper model that escalates more often can cost more overall.

05

Recommendations by volume tier

At low volume, use the strongest model and optimise later. At high volume, route routine tickets to a smaller model and escalate the rest.