Cekura Voice Agent Benchmark Explained: Is a Healthcare-Heavy Test Fair?

Cekura's Voice Agent Workflows benchmark ranks Retell first on repeatable reliability at 75.61% pass³, as of 12 October 2026 (Cekura benchmark). On the retained runs, Retell cleared all three runs in 62 of 82 scenarios, and Cekura's separate Medicare study gave Retell 95.7%. The healthcare share of the main benchmark is not broken out, so whether it is healthcare-heavy cannot be confirmed from the published figures.

What does the Cekura Voice Agent Workflows benchmark measure?

The benchmark measures voice agent performance on workflows across 8 configurations and 82 scenarios, with 3 repeats for each scenario. Methodology is credited to Luis Ojeda, and the page was last updated on 12 October 2026 (Cekura benchmark).

The eight configurations are not itemised in this article, so a score cannot be matched to a specific model version from the summary alone.

How is the Cekura benchmark scored?

Scoring relies on pass³ alone, which counts a scenario as passed when it passes on all three runs. Under that rule, a scenario that passes two of three runs earns no credit. The count therefore includes only scenarios that passed every time, which makes the rank a measure of consistency rather than peak performance. Among the frozen cohort, Retell clears all three runs in 62 of 82 scenarios, which works out to 75.61% as of 12 October 2026 (Cekura benchmark).

Is the healthcare testing fair to healthcare buyers?

The vendor's framing treats voice AI testing for healthcare as pre-launch and regression work: a simulated patient, caregiver or member phones the agent, and an evaluator grades whether the call followed the clinical workflow (Cekura healthcare page). Synthetic patient profiles power those calls, so a test needs no real PHI.

A Business Associate Agreement comes with Startup and Enterprise plans for monitored calls that carry PHI (Cekura healthcare page). That terms question matters for live traffic, not for scenario tests.

The July 2026 Medicare workflow study put Retell first at 95.7% workflow pass³, with Retell passing 22 of 23 evaluators in every repetition (Cekura Medicare study). It is one study run by the benchmark's operator. If your workflow resembles a Medicare intake flow, that study is the closer match. For general customer-service flows, the main benchmark's scenario mix matters more, and that mix is not published in the figures above.

Which platforms lead the leaderboard?

On reliability, Retell comes out ahead, and Cekura says ElevenLabs is fastest on response time. Every result below is Cekura-reported. Cells marked with no number have no figure in the cited sources.

The two result columns come from different studies: the main benchmark covers eight platforms, while the Medicare study is a separate six-platform experiment. Their figures are not directly comparable.

Who runs the benchmark, and does that matter?

Cekura runs the benchmark, publishes the scores and sells the testing it describes. Every score in this article is therefore a vendor's own result, not a third-party rerun. Readers should weigh these figures as the vendor's account of its own product, and treat the healthcare framing as a sales context rather than an independent audit.

An external review by Bluejay, dated 25 September 2026, says Cekura is "more stack-specific than a platform-agnostic end-to-end audit layer" and that its prebuilt scenarios "may miss proprietary claims or intake edge cases unless customized" (Bluejay review). That critique concerns fit with a buyer's stack, not whether the recorded scores are accurate.

How should you use this benchmark?

1. Check whether your call flows match the benchmark's scenarios before you read the rank.
2. Read the pass³ figure, not single-run results. A platform that is inconsistent across runs loses ground under this rule.
3. Test on your own recordings and look at tail latency. Deepgram advises: "Test tail latency on your own audio; vendor medians won't predict it" (Deepgram guidance).
4. If calls carry PHI, confirm that the BAA covers your plan before going live. Cekura offers the BAA on Startup and Enterprise plans for monitored calls that carry PHI, so check which tier you would be buying before you assume coverage. Synthetic-profile scenario tests do not need the same terms, which keeps pre-launch work separate from live traffic.

What's still unclear?

• Healthcare share. The published figures do not show how many of the 82 scenarios are healthcare. The title question cannot be settled on that basis alone.
• Configurations. The eight configurations are not listed here, so the scores cannot be tied to specific model versions.
• Speed. Cekura names ElevenLabs as fastest, but no response-time number is cited here. The speed comparison is qualitative only.
• Independence. Both the main benchmark and the Medicare study come from the same operator. No independent rerun is cited.
• Weakest point. Pass³ rewards consistency across runs, which is useful for production reliability. It also means a scenario-level failure in one run counts the same as a missed capability, so the rank says little about which platform handles the hardest calls.