Sonic 3.6 is the strongest option the public boards currently point to for real-time conversational speech — and the weakest fit among the top tier for audiobooks and long narration. Both halves of that sentence trace to the same independent evidence.
What it is
Sonic 3.6 is Cartesia’s streaming text-to-speech model, shipped August 18, 2026, and marketed primarily on speed. Cartesia’s stated latency figures place it in the “sub-90ms” class, aimed squarely at voice agents and other sub-second response loops. MarkTechPost covers the launch, but largely on vendor terms — its own caveat is that latency figures are vendor-stated and the honest move is to “benchmark your own round trip.” Sonic 3.6 was the standout of our August new-models roundup.
Coverage of 3.6 specifically is thin. Most independent measurement still targets Sonic 3 or Sonic 3.5, and the Vapi Humanness Index has no 3.6 entry yet. That matters when reading any score: much of what follows measures the previous generation, and a 3.x improvement should not be assumed to carry forward unexamined.
Where it sits across the boards
The aggregate picture is real but narrower than a headline suggests.
Artificial Analysis Speech Arena. Sonic 3.6 sits at #1 with an Elo around 1,288, ahead of Simba 3.2 at roughly 1,243. Model page / provider leaderboard. This is the stronger of the two Elo boards that currently measure 3.6, and it is the main driver of Sonic 3.6’s #2 spot on the texttospeech.com aggregate.
Vapi Humanness Index. No 3.6 entry yet. The closest entry, Sonic 3.5, sits around #11 in the Index. Model page. This is the gap in the record: the board that measures “humanness” is still looking at the previous version.
Voice Arena. Sonic 3.6 leads the US-English table at an Elo around 1,085. Voice Arena
So Sonic 3.6’s #2 spot on the aggregate is the output of our methodology. It tops both Elo boards that measure it, but it has no Vapi Humanness Index entry yet, and that missing third board caps its coverage, which is what keeps it behind Simba 3.2.
What wider independent evidence shows
The speed claim is where independent evidence diverges from marketing.
Coval, an independent voice-agent evaluation from an ex-Waymo team, measured Sonic-3 at a P50 latency of 188ms against Cartesia’s marketed ~90ms, with a latency spread roughly 50× wider than the best performer in the same test, alongside word error rates around 1.19% (Spanish) and 2.74% (Portuguese). Blog / benchmarks / code. Dated May 2026, so this is pre-3.6 — but the gap between vendor figure and measurement is the point.
Openbenchmarks, mirroring Coval and dated 2026-08-31, reports Sonic 3.6 at a 5.5% WER versus Eleven v3 at 4.4%, and Sonic 3.5 at a median TTFA of 271ms versus ElevenLabs at 187ms. Openbenchmarks. Sonic is fast; it is not the fastest measured.
The counterweight: Ry Walker’s independent testing puts Sonic 3.5’s real first-audio at 75–90ms over WebSocket — close to the claim when you measure first byte of audio on a live socket rather than a full round trip. Ry Walker. CloudX’s 30-plus stack benchmark measured Sonic-Turbo at 0.38–0.41s English TTFB. dev.to
The through-line: Sonic’s latency is real and competitive, but “sub-90ms” is a first-audio figure measured on a specific path. Independent round-trip and TTFA numbers land between 128ms and 271ms depending on path and version. Anyone quoting 90ms without naming what it measures is quoting the vendor.
Strengths
- Top of both Elo boards that measure it. #1 on Artificial Analysis Speech Arena at ~1,288 Elo and #1 on Voice Arena at ~1,085.
- Genuinely fast first audio. 75–90ms over WebSocket in independent testing, competitive across every independent latency benchmark we aggregate.
- Strong conversational quality. High blind-preference Elo, and a solid (if not top-ten) Vapi Humanness score for the 3.5 generation.
Weaknesses
- Long-form and audiobook output. Ry Walker notes the “texture difference becomes audible” on long-form, with “kinks on long utterances, random sounds.” Ry Walker. Practitioner sentiment on Hacker News says the same — “great quality, very fast — but with kinks (off the rails on long utterances, random sounds).” 2024 / 2024. TextToLab, an affiliate source and therefore weak, sums it up as “good enough for conversational agents but falls short for audiobooks.” TextToLab
- A small voice library. Roughly 100 voices versus 1,000+ at ElevenLabs, with less expressive cloning and no pronunciation dictionary or deep SSML — from Future AGI, a commercial source with its own interest in the comparison. Future AGI
- Latency marketing outruns measurement. The sub-90ms figure does not survive independent round-trip testing.
- Measured WER lags the leader on the Openbenchmarks run (5.5% vs 4.4% for Eleven v3).
Who it’s for
Real-time voice agents, IVR, live conversational interfaces, and low-latency demos. If the bottleneck is the sub-second loop between user speech and assistant speech, Sonic 3.6 is the strongest option the boards currently point to.
Who it’s not for
Audiobook production, long narration, expressive character work, or anything that needs a large, deeply controllable voice library. For those, the independent evidence consistently points elsewhere.
Bottom line
Sonic 3.6 is the leader for conversational latency and blind-preference quality, and its position on the texttospeech.com aggregate is defensible on that basis. Its weaknesses are specific and well-documented: long-form texture and a thin voice library. Choose it for live voice; choose something else for long-form narration.