Claude Haiku overtook ElevenLabs as the most variable stage in Loquent's phone pipeline

Vox-bench times each pipeline stage on its own and in combination. It fires an alert when any one stage's p95 crosses a configured threshold ^[3]. That split is the design choice worth copying. In June the team built its thresholds around TTS, which then had a 2.69x p99/p50 ratio ^[5]. A single round-trip number would have shown drift over the next four months. It could not have shown ElevenLabs getting steadier while Claude Haiku's tail widened over the same period ^[6].

The published figures check out against each other. Claude's 525ms p99 over a 185ms median is 2.84x, and June's 480ms over 210ms is 2.29x ^[2]. Dividing ElevenLabs' p99 by its ratio puts its median near 130ms in June and near 133ms now, so the Turbo v2.5 gain is almost entirely in the tail ^[1]. Claude moved in opposite directions at the two ends: a faster median and a slower p99 ^[7]. The team wrote that it thinks the change correlates with model updates that improved average performance but added variance under load ^[7].

Those numbers matter against the 680ms wall. The stage medians in the post are 175ms for Deepgram, 185ms for Claude and about 133ms for ElevenLabs. Together they come to roughly 493ms, leaving about 187ms for both Twilio legs and everything else in the turn ^[4]. Claude's p99 alone is 525ms, or 77% of the wall ^[3]. A turn with median STT, median TTS and a p99 LLM response comes to about 833ms before transport ^[5]. Percentiles do not add, so 833ms describes one bad turn shape. It is not an estimate of the pipeline's p99. According to the post, one bad turn is enough: callers who cross the wall start saying "hello?" or repeating themselves, and the call degrades even when later responses are fast ^[11].

The wall itself is a correlation from one caller base. It comes from per-call latency matched against completion across more than 40,000 calls over eight months of Loquent's Canadian healthcare and dental traffic ^[10][2]. The team has already moved it once. "We published our original article saying the threshold was 800ms. We were wrong," the team wrote ^[12]. The correction was 120ms, about 18% of the new figure ^[8].

The schedule has limits of its own. A GitHub Actions job every six hours gives four samples a day at fixed clock times ^[6]. A two-hour pattern like ElevenLabs' 2-4pm ET spikes in June shows up only when a run lands inside it ^[8]. The post does not report how many p95 alerts fired in four months or what they caught.

The evidence also supports keeping vendor comparisons running. Loquent runs its provider comparison mode continuously, putting the same workload through several vendors at once ^[15]. That mode caught an accent problem a latency threshold would miss. Google Cloud Speech-to-Text v2 ran a 195ms p50 against Deepgram's 175ms, but its word error rate was 18% higher on calls from patients in Quebec and Atlantic Canada ^[14][13]. Loquent stayed on Deepgram Nova-3 ^[13].