Ming-3.0-tts vs Cartesia Sonic 3.6: One Reads Your Invoice Number Right, the Other Answers in 90 ms

Put Ming-3.0-tts and Cartesia Sonic 3.6 on the same line of the Artificial Analysis speech arena and the scoreboard says Cartesia wins: 1,290 Elo against 1,188 on Provider Voice, a gap of 102 points, with the confidence intervals nowhere near touching. Put them on the pronunciation board and the result flips so hard it looks like a data error: 93.2% against 74.5%. The same two systems, measured by the same lab, in the same month, disagreeing by eighteen percentage points about which one can read English aloud. Both numbers are correct, and the reason they can both be correct is the whole comparison. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Ming-3.0-tts is Ant Group's MIT-licensed open-weights model — a 0.5B dense checkpoint you run on your own GPU, with a 7.9B mixture-of-experts sibling called Ming-3.0-tts-tiny. Cartesia Sonic 3.6 is a proprietary real-time endpoint released on 17 August 2026, priced per character, marketed on first-audio latency under 90 milliseconds and coverage of 44 languages. One is a model. The other is a service built to answer a phone call. Which one you want depends on a question that has nothing to do with Elo: is your text mostly prose, or mostly things that must be said exactly?

The two numbers that matter, and why they disagree

Start with the preference boards, because they are the ones people quote.

• Provider Voice Elo — Ming-3.0-tts 1,188 (interval 1,174–1,202) vs Sonic 3.6 1,290 (1,276–1,304). Sonic is 4th of 97 models; Ming is 18th.

• Controlled Voice Elo — Ming-3.0-tts 1,068 (1,052–1,084) vs Sonic 3.6 1,151 (1,138–1,164). Sonic is 5th; Ming is 12th. The gap narrows by 19 points once both systems clone the same eight voices.

• Pronunciation Robustness, English — Ming-3.0-tts 93.2% (91.5–94.9) vs Sonic 3.6 74.5% (71.4–77.5). Ming is 1st of 31 measured systems; Sonic is 19th.

Those are not contradictory results; they are measuring different things. The Elo boards ask human listeners which clip they prefer, and preference rewards the things that make synthetic speech pleasant — natural prosody, convincing emotion, a voice that does not sound like it is reading a manual. Sonic 3.6 is very good at that. It has been the fourth-ranked system in the arena since August and it earns the position.

The pronunciation board asks a narrower, harsher question: did the system say this specific span correctly? Human reviewers marked up highlighted words and phrases in English text — brand names, technical terms, file paths, email addresses, numbers, dates, units — and judged each one. On that test, Sonic lands in the bottom half of the 31 systems that have been scored, and Ming lands at the top of all of them.

The sub-scores make the shape of the failure clearer than the aggregate does:

• Standalone Terms — Ming-3.0-tts 88.5% vs Sonic 3.6 72.6%

• Contextually Appropriate — Ming 91.0% vs Sonic 78.2%

• Preserving Exact Sequences — Ming 90.6% (1st of 31) vs Sonic 60.8%

• Expanding Shorthand — Ming 96.5% (1st of 31) vs Sonic 49.1%

Expanding Shorthand is the one to stare at. It covers numbers, dates, units and notation — "6'2"", "Chapter XVII", "1 tsp". Sonic 3.6 gets roughly half of those spans right by this measure. That is a startling number for a production voice API, and it is the kind of thing a demo never surfaces because demos are written in prose that the vendor chose. Put an order confirmation, a dosage, or a measurement in front of it and the preference ranking stops predicting anything useful.

The latency and throughput case, where Cartesia is not close

Cartesia's headline claim is first audio in under 90 milliseconds, which is the number that matters for a conversational agent: the caller hears a voice before they decide you are a machine. Ming-3.0-tts publishes no latency figure at all. What the independent harness has is throughput, and it is not kind to the open model.

• Characters per second of generation — Sonic 3.6 146.9 vs Ming-3.0-tts 44.3. Cartesia is 3.3 times faster on identical hardware conditions.

• Median generation time for a scored request — Sonic 3.6 9.20 s vs Ming 2.56 s. Read this one carefully; it reflects different request lengths, not a speed advantage for Ming.

The honest summary is that Sonic 3.6 is built for streaming and Ming-3.0-tts is built for batch. At 44.3 characters per second — roughly seven words — the self-hosted model does outrun real speech, so a long narration finishes ahead of the time it takes to listen to it. What it does not have is a streaming story, a published time-to-first-audio, or session handling. Cartesia sells exactly those three things.

If you are building a phone agent, a live translation tool, or anything where a human is waiting on the first syllable, that is the end of the comparison and you should stop reading. The rest of this piece is for the other case: text that is prepared in advance, or text whose correctness matters more than its immediacy.

What each side actually gives you

Ming-3.0-tts, from Ant Group's inclusionAI group, is distributed as a 2.96 GB checkpoint containing the backbone, a diffusion-transformer patch head, an audio tokenizer, a speaker encoder and a text tokenizer, under the MIT licence with commercial use permitted. It outputs 44.1 kHz audio, supports seven languages plus Sichuanese and Cantonese, and exposes voice cloning, text-described voice design, 14 inline paralinguistic tags, emotion and speed/pitch/volume control. Ant tested it on an H800-80GB and an H20-96G with PyTorch 2.6 and CUDA 12.4, and ships a Dockerfile.

Sonic 3.6, from Cartesia, is an API. It handles 44 languages from one model, clones a voice from ten seconds of reference audio, supports emotion-aware delivery and insertable non-verbal cues, and accepts custom pronunciation dictionaries — which is the vendor's own answer to exactly the problem the pronunciation board measures, and worth noting that a customer who configures it is not being scored on the default path the arena tests. Cartesia also builds the rest of the voice stack: speech-to-text, voice agents at $0.06 per minute, telephony numbers, and concurrency limits that scale with plan tier.

• Licence — MIT weights you keep vs a commercial API with no weights

• Language coverage — 7 plus 2 dialects vs 44

• Deployment — your GPU vs their endpoint

• Output sample rate — 44.1 kHz vs not published on the product page

• Voice cloning — reference audio, reference text and an explicit task mode vs 10 seconds of audio, instant

• Pricing basis — GPU hours vs credits at one credit per character

The credit pricing is worth converting, because "one credit per character" is not a number most teams can budget against. Cartesia's own plan table says a minute of audio takes 750–800 credits, which puts the effective rate near $0.038 per minute on the $5 Pro plan, $0.029 on the $49 Startup plan and $0.028 on the $299 Scale plan. The arena normalises it to $49 per million characters, which is the comparable figure against Ming's zero. Ming is not free; it is a GPU-hour bill that your finance team will see as infrastructure rather than as a line item per word.

The three-month comparison nobody puts in the review

There is a version of this decision that outlives both models, and it is the one that matters for a production system: what happens when the model you picked is replaced.

Cartesia has shipped Sonic 3, 3.5 and 3.6 within a year, at a constant $49 per million characters. Ant has shipped Ming-omni-tts in February, a 16.8B variant in June, and now Ming-3.0-tts — with a technical report still unchecked on its roadmap and a vLLM-Omni acceleration path marked done but undocumented in the repository. Both vendors move quickly. Neither is going to stop.

For a hosted model that is a routine version bump. For a self-hosted checkpoint it is a migration: new weights, new dependencies, a new round of evaluation on your own text. The MIT licence protects your right to keep running the old one, which is a real advantage and the main reason teams choose open weights in the first place — but it does not protect you from wanting the new one.

This is the seam where a routing layer does something concrete rather than rhetorical. OrcaRouter does not host Ming-3.0-tts or Cartesia Sonic 3.6, and nothing here should suggest otherwise — Ming comes from the repository, Sonic from Cartesia's own API. What the catalogue does carry is a set of hosted speech models, including G​oogle's G​emini T​TS previews and O​penAI's tts-1 and gpt-4o-mini-tts, behind the same key as more than 200 text models, at provider list price with no markup added, with automatic failover between providers. If your voice pipeline has a fallback path — and any pipeline that handles live traffic should — that fallback can be one endpoint rather than a second contract and a second SDK. A vendor price cut lands on our side the same day, because there is no markup in between to unwind.

Which one to pick, by workload

The decision is short if you answer honestly about the text.

Pick Sonic 3.6 if a human is waiting on the first audio, you need more than a handful of languages, or you want a vendor to own the uptime. The 102-point preference gap is real and listeners will hear it. Configure the pronunciation dictionary for the terms your product actually says, because the default path scores 74.5% on exactly that.

Pick Ming-3.0-tts if your audio is generated ahead of time — narration, documentation read-aloud, IVR prompts, training material, audiobook drafts — and the text contains identifiers, units, codes or notation. On the three categories that cover those spans it is either first or within three points of it, and no hosted system on the board is better at preserving an exact sequence.

Pick neither, yet, if you need English proper nouns to be right. Ming's 88.5% on Standalone Terms is 21st of 31, and it is the one category where the model is unambiguously mid-table. A brand name mispronounced in a product demo costs more than a date read awkwardly.

There is also a hybrid that neither vendor will suggest: use a hosted real-time model for the interactive surface and a self-hosted checkpoint for the batch corpus, and evaluate them on the same held-out text. The two systems fail in different places, which is what makes the split worth the operational cost.

The open question

Cartesia's product page says the company is ranked first on the Speech Arena. That was true when the page was written and is no longer true — HeyGen Voice, Q​wen-Audio-3.1-T​TS-Plus, Eleven v4 Turbo and Eleven v4 all sit above Sonic 3.6 today, and Sonic's own position is fourth. Vendor leaderboard claims age in weeks. The 74.5% pronunciation score is the more durable fact in this comparison, because it is measured by someone with no product to sell and it is bad enough that Cartesia's own pronunciation-dictionary feature exists to work around it.

What would change the verdict: a streaming release of Ming-3.0-tts with a published time-to-first-audio, or a Sonic 3.7 that fixes the shorthand expansion. Neither is announced. Until then the split is clean — speed and coverage on one side, exactness and ownership on the other — and the right answer is to stop looking for a winner and start looking at your own corpus.