Smallest AI vs Cartesia

Most comparisons between these two start with latency, and that is the wrong place to start. Both are fast enough. Cartesia quotes sub-90ms time-to-first-audio for Sonic-3.6 and Lightning TTS quotes sub-100ms, and once you put either behind a real network the measured numbers converge into a band where the difference stops mattering to the person on the phone. What actually separates them is what happens after the audio.

Cartesia is a text-to-speech company that has since added speech recognition and a voice agent product. Smallest.ai is a voice agent platform where the speech models happen to be its own. That single architectural difference explains almost every trade-off in this comparison, and it is the thing to hold on to while reading the rest.

Before the detail.

• Cartesia leads on benchmark naturalness. Sonic-3.6 holds the top position on both Artificial Analysis speech leaderboards as of August 2026. If your product is judged on how good a single voice sounds in isolation, that is a real and independently verifiable advantage.

• Smallest.ai leads on the full pipeline. Pulse STT, Lightning TTS, the Electron speech-to-speech model, and the Atoms SDK come from one vendor under one contract, which collapses the three-vendor stitching most voice agent teams end up doing.

• Language coverage splits by geography. Cartesia covers 44 languages and 61 locales broadly. Lightning covers 70+ languages, accents, and dialects with unusual depth on Indic languages and mid-sentence code-mixing, which matters enormously if your users are in India or the diaspora.

• Pricing models are not comparable as published. Cartesia sells credits where one credit is roughly one character. Smallest.ai sells per-minute agent time with the model layers itemised. Comparing the headline numbers directly produces a wrong answer.

• On-premise is the clearest dividing line. Smallest.ai productises on-premise deployment for regulated workloads. Cartesia does not.

Voice quality and naturalness

This is the section where an honest comparison has to concede something, so here it is. Cartesia released Sonic-3.6 as the newest version of its real-time text-to-speech model, roughly three months after Sonic-3.5, and the headline change is naturalness. Sonic 3.6 holds the top position on both Artificial Analysis speech leaderboards, with 1,283 Elo on the Provider Voice board and 1,123 on the Controlled Voice board. The second number is the one that carries weight. That board clones every model onto the same eight reference voices, which isolates the synthesis engine from the voice catalog, so it measures the model rather than the library.

Lightning TTS does not currently hold that position. What it does deliver is broadcast-grade 44.1kHz output that holds up across long-form narration, and a prosody profile tuned for conversational turn-taking rather than for audiobook read-through. In practice the gap between the two on a 30-second phone call is far narrower than the gap on a 30-minute audiobook, because conversational audio is dominated by short utterances where expressive range has less room to show itself.

Trade-off worth naming. If your product is an audiobook platform, a podcast tool, or anything where a listener sits with one voice for an extended stretch, Cartesia's benchmark lead is worth taking seriously and you should A/B the two on your own scripts. If your product is a voice agent handling three-minute support calls, the quality difference is unlikely to be the thing that decides your architecture.

Latency, and why the published numbers mislead

Cartesia's marketing leads with sub-90ms time-to-first-audio. Lightning TTS leads with sub-100ms. Both are model-latency figures measured under favourable conditions, and neither includes your network.

Independent measurement puts this in perspective. Real-world latency for Sonic measures around 166 to 190 milliseconds, and the sub-90ms figure is a vendor model-latency claim under ideal conditions, with the two numbers never to be conflated. The same caution applies to every vendor in this category including Smallest.ai. What you should plan against is the real-world band, not the press-release number.

The more useful question for a voice agent is what the whole turn costs, not what the TTS layer costs. A conversational turn is speech recognition plus language model plus synthesis plus network on both ends. Smallest.ai publishes 64ms time-to-first-transcript for Pulse STT and 175ms synthesis for Lightning TTS, which combine toward total turn times under 800 milliseconds when the language model is fast. Assembling the equivalent from Cartesia means pairing Sonic with either Cartesia's own Ink STT or a third-party recognition layer, and each additional vendor hop adds network round-trips that do not show up in anybody's model-latency benchmark.

Trade-off worth naming. Cartesia's Line product does bundle a voice agent stack, so the single-vendor argument is not exclusive to Smallest.ai. The difference is depth. Smallest.ai's Atoms SDK exposes session and node primitives for multi-node orchestration, background compliance logging, and warm or cold human transfer as first-class concepts, which is a more complete agent framework than a bundled endpoint.

Language and accent coverage

Cartesia's current model, Sonic-3.6, covers 44 languages across 61 locales, up from 42 languages in the previous generation, with recent additions including Odia, Urdu, and Hinglish. Cartesia describes step-change multilingual performance particularly in Hebrew, Japanese, Spanish, Hindi, German, Korean, and French. That is genuinely broad coverage and the recent Indic additions narrow a gap that used to be wide.

Lightning TTS covers 70+ languages, accents, and dialects, with dedicated depth across Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi, and Odia. The differentiator is not the count but the handling of code-mixing. Lightning does automatic language detection with seamless code-mixing mid-sentence, which is the actual speech pattern of most Indian users. A caller who says "mera order kahan hai, the tracking says delivered" is not switching languages between turns, they are switching inside a clause, and a model that requires the language to be declared at session start will degrade at that boundary.

Trade-off worth naming. For a product serving primarily European, North American, or East Asian markets, Cartesia's locale breadth is likely the better fit and the naturalness lead compounds. For a product serving India, the Gulf, or South Asian diaspora markets, Lightning's code-mixing behaviour is the more consequential capability and worth testing on your own recordings before deciding.

Emotional control and expressiveness

Cartesia's approach centres on inline markup. Sonic supports inline nonverbal tags like laughter, and the model reads alphanumerics correctly for confirmation codes, order numbers, phone numbers, and email addresses, which is precisely where voice agents used to embarrass themselves. It also resolves English heteronyms such as read, bass, and bow from context. Those are practical wins for transactional voice work and worth acknowledging plainly.

Lightning's expressiveness is tuned toward emotional range within a consistent speaker identity, which is the requirement for a support agent that needs to sound calm on turn one and empathetic on turn twelve without the voice appearing to change person. For gaming and character work, the model exposes dynamic emotional range across character voices.

Where this actually lands. Both platforms have moved past the era where you had to hand-tune SSML for every sentence. Cartesia gives you finer explicit control through tags. Lightning leans more on inferring the right delivery from context. Which you prefer is closer to a workflow preference than a capability gap, and teams that want deterministic output usually prefer tags while teams that want less prompt engineering usually prefer inference.

Voice cloning

Both clone from roughly ten seconds of audio with no studio equipment. The meaningful difference is what cloning costs you at scale.

Cartesia's professional voice cloning changes the billing rate. Cartesia charges 1 credit per character for standard TTS, but if you use Pro Voice Cloning the rate jumps to 1.5 credits per character, which is a 50% effective price increase on every character generated with a cloned voice. For a product built around cloned voices rather than catalog voices, that multiplier is the single most important line in Cartesia's pricing and it is easy to miss.

Smallest.ai's instant voice cloning produces a production-ready clone in under ten seconds and does not apply a separate per-character multiplier to cloned-voice generation.

Pricing, compared honestly

The two pricing models are structurally different, which is why most comparisons of these platforms get this section wrong.

Cartesia uses a credit system for TTS where roughly one credit equals one character, plus separate voice agent usage, with plans ranging from Free at 20,000 credits per month to Scale at 8 million credits per month. The Pro plan is $4 per month on annual billing or $5 per month on monthly billing. Scale at $299 per month includes roughly 10,667 TTS minutes and 15 concurrent requests, and Line voice agents bill separately at $0.06 per minute. The effective cost works out to roughly $5 to $37 per million characters depending on plan.

Smallest.ai prices agent time per minute with the model layers itemised, so you can see what each component costs.

The comparison that matters is total cost per conversation minute, not cost per character. A voice agent generating roughly 150 spoken words per minute produces something near 900 characters of TTS output per minute. On Cartesia's Scale plan at 8 million credits for $299, that is roughly 8,900 minutes of synthesis, or about $0.034 per minute for TTS alone, before adding Line agent minutes at $0.06 and before any speech recognition. On Smallest.ai the equivalent all-in agent minute is published at $0.09 to $0.21 with everything included.

Read that carefully rather than as a scoreboard. Cartesia can come out cheaper on pure synthesis volume, particularly on the Scale tier, because credits bought in bulk are inexpensive per character. Smallest.ai comes out more predictable on full agent workloads because the per-minute rate already contains the recognition, hosting, and concurrency that Cartesia bills separately or not at all. If you are generating audio files, model Cartesia. If you are running conversations, model per-minute.

One more asymmetry worth knowing. Cartesia's Scale plan includes 15 concurrent requests. Smallest.ai includes 20 concurrent streams on pay-as-you-go. For a call centre replacement, concurrency limits bind before price does, and discovering that at launch is an expensive surprise.

Deployment, compliance, and the on-premise question

Both platforms carry serious compliance posture. Cartesia holds HIPAA and SOC 2 Type 2 attestation. Smallest.ai holds ISO 27001, SOC 2 Type 2, GDPR, and HIPAA.

The divergence is deployment topology. Smallest.ai productises on-premise deployment, including air-gapped environments, for healthcare, banking, and government workloads, and lists it as an Enterprise plan feature alongside 99.99% uptime SLA and HIPAA zero-data-retention. Cartesia does not offer a comparable productised on-premise path.

For a regulated enterprise where audio cannot leave the network perimeter, this is not a preference, it is a filter. The shortlist reduces to vendors that ship on-premise, and Cartesia is not currently on it regardless of how well Sonic-3.6 scores on naturalness.

The API surface

Lightning TTS via a single POST.

const res = await fetch(
  "https://api.smallest.ai/waves/v1/lightning-v3.1/get_speech",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer YOUR_API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      text: "Modern problems require modern solutions.",
      voice_id: "magnus",
      sample_rate: 44100,
      output_format: "wav",
    }),
  },
);

writeFileSync("output.wav", Buffer.from(await res.arrayBuffer()));

const res = await fetch(
  "https://api.smallest.ai/waves/v1/lightning-v3.1/get_speech",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer YOUR_API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      text: "Modern problems require modern solutions.",
      voice_id: "magnus",
      sample_rate: 44100,
      output_format: "wav",
    }),
  },
);

writeFileSync("output.wav", Buffer.from(await res.arrayBuffer()));

const res = await fetch(
  "https://api.smallest.ai/waves/v1/lightning-v3.1/get_speech",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer YOUR_API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      text: "Modern problems require modern solutions.",
      voice_id: "magnus",
      sample_rate: 44100,
      output_format: "wav",
    }),
  },
);

writeFileSync("output.wav", Buffer.from(await res.arrayBuffer()));

Cartesia Sonic via its Python SDK.

import os
from cartesia import Cartesia

client = Cartesia(api_key=os.environ["CARTESIA_API_KEY"])

data = client.tts.bytes(
    model_id="sonic-3.6",
    transcript="Hello, world!",
    voice_id="your-voice-id",
    output_format={
        "container": "wav",
        "encoding": "pcm_f32le",
        "sample_rate": 44100,
    },
)

with open("cartesia.wav", "wb") as f:
    f.write(data)

import os
from cartesia import Cartesia

client = Cartesia(api_key=os.environ["CARTESIA_API_KEY"])

data = client.tts.bytes(
    model_id="sonic-3.6",
    transcript="Hello, world!",
    voice_id="your-voice-id",
    output_format={
        "container": "wav",
        "encoding": "pcm_f32le",
        "sample_rate": 44100,
    },
)

with open("cartesia.wav", "wb") as f:
    f.write(data)

import os
from cartesia import Cartesia

client = Cartesia(api_key=os.environ["CARTESIA_API_KEY"])

data = client.tts.bytes(
    model_id="sonic-3.6",
    transcript="Hello, world!",
    voice_id="your-voice-id",
    output_format={
        "container": "wav",
        "encoding": "pcm_f32le",
        "sample_rate": 44100,
    },
)

with open("cartesia.wav", "wb") as f:
    f.write(data)

Both are clean. Neither will be the reason you pick a platform. The more meaningful developer-experience difference sits one layer up, where Smallest.ai's Atoms SDK gives you AgentSession, OutputAgentNode, and BackgroundAgentNode primitives for building the orchestration layer, rather than leaving you to assemble it around a synthesis endpoint.

Which to pick

Pick Cartesia if your product is judged on single-voice naturalness in long-form audio, your markets are primarily European, North American, or East Asian, your workload is high-volume synthesis rather than live conversation, and cloud deployment is acceptable.

Pick Smallest.ai if you are building voice agents rather than generating audio files, your users speak Indic languages or code-mix mid-sentence, you want speech recognition, synthesis, and orchestration from one vendor under one contract, you need predictable per-minute economics with concurrency included, or you have an on-premise requirement.

Run both if you are early enough that the decision is reversible. Generate the same 20 utterances from your actual product script on both platforms, play them to five people who are not on your team, and measure which one they prefer. Vendor benchmarks are run on vendor-chosen audio, and neither Artificial Analysis nor any marketing page knows what your callers sound like.

For a deeper look at Cartesia specifically, see our Cartesia AI review covering features and pricing and our roundup of Cartesia alternatives for text-to-speech.

Start with Lightning TTS on $10 of free credits, or talk to the team if your evaluation involves on-premise or regulated workloads.