ElevenLabs eleven_v3 production run: 271 calls, $0.002 each, no failures

What happened

A usage log published on dev.to by a developer at the one-person studio NeuraGrowth gives a rare look at ElevenLabs' eleven_v3 speech model doing real work. Over ten days, from October 28 to November 6, 2026, the model handled 271 requests from a single pipeline job, kids_episode.tts, which turns episode scripts into audio. According to the post, not one call failed.

The headline figures from the log:

• 271 calls, all successful
• $0.5716 in total spend, about $0.00211 per call
• A median response time of 1.4 seconds and a worst case of 4.1 seconds

Speech was the cheap part of the pipeline

The same ten-day window covered 213 calls to claude-opus-5-5, the language model doing the writing and research that produced the scripts in the first place. Those calls cost $25.4973, roughly $0.12 each — almost sixty times the per-call cost of the speech step. In the author's accounting, the TTS line item was nearly invisible next to the language-model work feeding it.

Latency was tight

Response times clustered closely: half of all calls finished within 1.4 seconds, and the slowest of the 271 took 4.1 seconds. The author highlights how narrow that spread is, which matters for anyone provisioning timeouts or capacity around a speech stage.

What the log does not show

The record has limits the author is careful to state. It does not capture how much text each request contained, so it is impossible to tell whether input length influenced the timing figures. And because every call came from one job over one ten-day stretch, the numbers describe this workload rather than eleven_v3 in general. Treating them as a universal benchmark would overread the data.

The author's operational advice

From these measurements, the author suggests two practices: set text-to-speech timeouts based on the median and worst-case times actually observed for the specific job, and keep TTS metrics separate from language-model metrics. The reasoning is that the two stages sit orders of magnitude apart in both cost and duration, so lumping them together can hide problems in the speech step.

Why it matters

Public production telemetry for AI voice APIs is scarce; most cost discussion relies on list prices rather than observed behaviour. A log like this gives voice-product builders a concrete reference point: a fraction of a cent per generation, zero failures across a few hundred calls, and a latency band tight enough to plan around. It also reframes budgeting for multimodal pipelines. If the script-writing model costs tens of dollars while rendering the finished speech costs cents, optimisation effort arguably belongs upstream of the audio stage. The usual caveats apply — a single workload, a modest sample, and unknown input lengths — but as an early cost and reliability signal for eleven_v3 in production, it is useful data for anyone building AI voice features.