We have spent the last six months running all three major streaming-TTS providers on a single production voice agent, switching weighted traffic between them and logging every metric we could think of. There is no overall winner. There is a clear winner per use case.
How we set up the test
Same agent, same prompts, same VAD settings, same TTS voice profile per provider (matched as closely as the voice catalogs allow). Traffic split 33/33/33 with a per-call random seed so the assignment was deterministic for QA replay.
We measured time-to-first-audio (TTFA), perceived naturalness via post-call survey, cost per minute of synthesized speech, and incident rate (timeouts, malformed audio, mispronunciations of named entities).
Latency: TTFA at p50 and p99
ElevenLabs sits around 110ms p50 / 220ms p99 in the same region. Cartesia is the fastest by a clear margin (75ms p50 / 160ms p99), and Sonic-2 is faster still on short utterances. OpenAI is the slowest in this test, around 180ms p50 / 380ms p99, with noticeable tail variance.
Naturalness
On long-form expressive content (storytelling, product walkthroughs), ElevenLabs still wins survey scores: its prosody is the most expressive of the three. Cartesia is close behind and noticeably more consistent across voices. OpenAI is fine, not great; it doesn't embarrass itself, but it rarely shines either.
Cost
Cartesia is cheapest by a wide margin on volume, especially after you negotiate. ElevenLabs is the most expensive but predictable. OpenAI sits in the middle.
Reliability
Across 1.2 million minutes synthesized, each provider had at least one regional incident. Cartesia had the fewest. ElevenLabs had the most graceful degradation when things did break: quality dropped before it failed outright.
Verdict
- Latency-critical voice agents: Cartesia.
- Expressive long-form: ElevenLabs.
- Multimodal apps already on OpenAI: OpenAI, mainly for the operational simplicity of one vendor.