Cartesia is making Sonic-3.6 generally available , its latest text-to-speech model for developers building voice agents, alongside its Ink-2 streaming transcription model. The company describes Sonic-3.6 as native in 44 languages, matching conversational tone, pacing, and emphasis , and positions the pair as a single-vendor stack for teams that currently stitch together separate STT and TTS providers.
The pitch to founders and operators shipping voice products is coherence across the pipeline. Cartesia claims the only STT and TTS optimized across the full real-time loop, one API for both models, and sub-90ms TTS with roughly 100ms transcript latency plus native turn detection . On third-party benchmarks, the company points to a #1 ranking on Artificial Analysis's Speech Arena and Speech-to-Text leaderboards , and the launch tweet frames this release as topping their own prior model rather than iterating on a competitor's baseline.
The technical story matters here because of who is behind it. Cartesia was co-founded by CEO Karan Goel with Albert Gu, Arjun Desai, Brandon Yang, and Chris Ré , the team that spent years developing the state space model (SSM) architecture that now underpins Sonic and Ink. The company argues the usual voice AI tradeoffs between speed and naturalness, or accuracy and cost, are architectural rather than inevitable, and that SSMs are the primitive that lets them collapse those constraints , which is the through-line connecting this launch to the research bet the founders made before starting the company.
Comments (8)
Sign in to join the discussion.
N
Nadia OkaforAug 28, 2026
Two months to beat your own SOTA is either extreme confidence or extreme regret about the last release. Either way I'm listening.
B
Boris PlekhanovAug 28, 2026
open weights? evals repo? or is 'purpose-built for voice agents' code for 'trust us bro'
T
Tomas L.Aug 28, 2026
ok wait, 'rebuild from the architecture up' in January and shipping GA now? that's the kind of timeline that makes me nervous about my own backlog.
R
Rasheed V.Aug 28, 2026
naive question but if it's real-time and also best quality, what's the actual tradeoff you removed? something has to give, right?
P
Priyanka ShahAug 28, 2026
the thumbnail slaps but that waveform gradient is doing a LOT of heavy lifting for one product shot. would've loved to actually hear a sample in the tweet itself.
K
kwame d.Aug 28, 2026
blog post title is doing marketing that the tweet copy refused to do. bury the lede much.
M
Min-ji FolauAug 28, 2026
how does pricing scale when a voice agent runs a 40 minute call? per-second billing gets scary fast for support use cases.
I
iolantheAug 28, 2026
the models get faster, the agents get chattier, the humans get quieter. we are training ourselves out of the conversation.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Cartesia
The #1 real-time speech and transcription models purpose-built for voice agents. One API, no tradeoffs between quality and speed.
The #1 real-time speech and transcription models purpose-built for voice agents. One API, no tradeoffs between quality and speed.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Cartesia is making Sonic-3.6 generally available , its latest text-to-speech model for developers building voice agents, alongside its Ink-2 streaming transcription model. The company describes Sonic-3.6 as native in 44 languages, matching conversational tone, pacing, and emphasis , and positions the pair as a single-vendor stack for teams that currently stitch together separate STT and TTS providers. The pitch to founders and operators shipping voice products is coherence across the pipeline. Cartesia claims the only STT and TTS optimized across the full real-time loop, one API for both models, and sub-90ms TTS with roughly 100ms transcript latency plus native turn detection . On third-party benchmarks, the company points to a #1 ranking on Artificial Analysis's Speech Arena and Speech-to-Text leaderboards , and the launch tweet frames this release as topping their own prior model rather than iterating on a competitor's baseline. The technical story matters here because of who is behind it. Cartesia was co-founded by CEO Karan Goel with Albert Gu, Arjun Desai, Brandon Yang, and Chris Ré , the team that spent years developing the state space model (SSM) architecture that now underpins Sonic and Ink. The company argues the usual voice AI tradeoffs between speed and naturalness, or accuracy and cost, are architectural rather than inevitable, and that SSMs are the primitive that lets them collapse those constraints , which is the through-line connecting this launch to the research bet the founders made before starting the company.
Comments (8)
Two months to beat your own SOTA is either extreme confidence or extreme regret about the last release. Either way I'm listening.
open weights? evals repo? or is 'purpose-built for voice agents' code for 'trust us bro'
ok wait, 'rebuild from the architecture up' in January and shipping GA now? that's the kind of timeline that makes me nervous about my own backlog.
naive question but if it's real-time and also best quality, what's the actual tradeoff you removed? something has to give, right?
the thumbnail slaps but that waveform gradient is doing a LOT of heavy lifting for one product shot. would've loved to actually hear a sample in the tweet itself.
blog post title is doing marketing that the tweet copy refused to do. bury the lede much.
how does pricing scale when a voice agent runs a 40 minute call? per-second billing gets scary fast for support use cases.
the models get faster, the agents get chattier, the humans get quieter. we are training ourselves out of the conversation.