Real-World Latency Numbers for GPT-5, Claude, and Gemini

Hardware is a moving target. The Jetson Orin we benchmarked in January was outperformed by an off-the-shelf mini-PC by August. We re-run the benchmark matrix every quarter and have stopped making long-term hardware commitments.

Background

Cost modelling is now part of our pre-merge checklist. Every PR that touches an LLM call includes an estimate of the per-request token spend and the expected daily volume. Surprises in the monthly invoice have dropped to nearly zero.

In production, latency distributions matter far more than averages. A pipeline whose mean response time looks acceptable can still feel sluggish if the 95th percentile drifts upward during peak hours. We instrument every stage with histograms so regressions surface immediately.

Results

The first version of this system was deliberately simple. We wanted a baseline that could be measured against, rather than an architecture that anticipated every possible failure mode. That decision paid off — most of the issues we eventually hit were unrelated to the ones we had originally feared.

When the system is wrong, the user should be able to understand why in under thirty seconds. Citation links, confidence scores, and the exact retrieved passages are surfaced in the UI for every generated answer.

Observability for agent runs is qualitatively different from traditional APM. A single user request can spawn dozens of tool calls, each with its own latency, cost, and failure mode. Flat traces become unreadable; we render them as collapsible trees.