public benchmark registry

Compare latency evidence, not one mystery number.

Published results grouped by what was actually measured: full voice turns, orchestration paths, or carrier network round trips. Sources checked 2026-07-25.

Keep unlike measurements in separate lanes

Lower is faster inside the same scope. Different prompts, models, carriers, regions, endpointing, and test owners can still change the result.

Published response latency

01

Retell AI

500 production calls per platform in March 2026

third party
680 mspublished typical / p50
920 msp95
source and caveat

Reproduce with your own carrier, models, prompts, tools, caller regions, and endpointing settings.

open Tested Media source ↗
02

Vapi optimized

500 production calls per platform in March 2026

third party
720 mspublished typical / p50
1,050 msp95
source and caveat

The result depends on the selected transcriber, model, voice, region, and tuning.

open Tested Media source ↗
03

Bland AI

500 production calls per platform in March 2026

third party
850 mspublished typical / p50
1,180 msp95
source and caveat

Campaign traffic, tool calls, concurrency, and destination routes can change the result.

open Tested Media source ↗
04

Synthflow

500 production calls per platform in March 2026

third party
920 mspublished typical / p50
1,250 msp95
source and caveat

The published result reflects the tested configuration, not every Synthflow deployment.

open Tested Media source ↗

Published response latency

01

Twilio ConversationRelay

Internal benchmark using different model configurations

official internal
491 mspublished typical / p50
713 msp95
source and caveat

The product page does not fully define the customer application, model path, or network boundaries.

open Twilio source ↗
02

SignalWire fastest config

Smart Appointment Assistant on a fixed model and tool stack

provider-authored comparison
1,090 mspublished typical / p50
not publishedp95
source and caveat

SignalWire authored and participated in this comparison. Treat it as reproducible evidence, not an independent award.

open SignalWire source ↗
03

SignalWire average

Average across 5 SignalWire configurations

provider-authored comparison
1,240 mspublished typical / p50
not publishedp95
source and caveat

The report compares configurations under SignalWire’s stated test design.

open SignalWire source ↗
04

LiveKit tuned

Smart Appointment Assistant on a fixed model and tool stack

provider-authored comparison
1,750 mspublished typical / p50
not publishedp95
source and caveat

SignalWire authored this result. Verify the configuration and rerun it before procurement.

open SignalWire source ↗

Network leg only

01

Telnyx

Independent carrier-leg network round-trip test cited by Telnyx

provider-authored comparison
118 mspublished typical / p50
not publishedp95
source and caveat

Carrier RTT measures only the network leg. It excludes endpointing, models, tools, synthesis, and playback.

open Telnyx comparison source ↗
02

Twilio

Independent carrier-leg network round-trip test cited by Telnyx

provider-authored comparison
161 mspublished typical / p50
not publishedp95
source and caveat

Carrier RTT is not comparable with a complete end-of-speech to first-audio result.

open Telnyx comparison source ↗
reproduce itUse the latency benchmark labto enter your own chronological trace. Procurement should use your carrier, caller region, production model, tools, and concurrency.