Sonic 3.5 vs Inworld TTS 2

Sonic 3.5 is the latency and technical-text row. Inworld TTS 2 is the character and style-steering row. Sonic 3.5 bills 4 credits per 1,000 characters on Versely and Inworld TTS 2 bills 2. Cartesia Sonic 3.5 is Cartesia’s previous flagship TTS (the newer Sonic 3.6 is the same 4 credits and backwards compatible): multilingual, expressive, emotion control, speed tuning, voice cloning. Inworld TTS 2 is Inworld’s research-preview flagship: 100+ languages with natural-language style steering for fine-grained delivery. Search Console still shows impressions for “sonic 3.5” and “inworld tts-2” (12 and 11 impressions, position 9, 7 days to 4 September 2026) without a pick page. This is the pick. It is not a latency explainer.

If the file must come back fast and read a SKU, a URL, or a phone number, start Sonic. If the file must stay one character across a script, start Inworld.

Two cheap voice files, not talking video

Both cards are audio. Content type audio. Categories include text-to-audio. They return a voice file. They do not lip-sync a face. They do not generate a talking clip. If you needed a mouth in shot, that is a video or lipsync row. Then you might still use these for a bed or a VO on a silent plate. Text-to-speech is the term. Do not open Veo because you needed a read.

Sonic 3.5 catalog. 4 credits per 1,000 characters. text-to-audio and voice-clone. Cartesia’s previous flagship TTS, multilingual, emotion control, speed tuning, cloning. Cartesia’s own model docs (cited in the latency post) put first-audio latency under 90 milliseconds and native support across 42 languages, with stronger handling of numbers, codes, emails, and heteronyms. That spec is why Sonic shows up in live-agent stacks. For rendered voiceover, the viewer never hears the 90 milliseconds. What they hear is whether “SKU-104B” and “record” (noun vs verb) landed. Sonic is the row when those strings are the job.

Inworld TTS 2 catalog. 2 credits per 1,000 characters. text-to-audio. Research preview. Most powerful and expressive Inworld model. 100+ languages. Natural-language style steering. Inworld’s lineage is character: NPCs that stay the same person across many lines, with anger, sarcasm, or quiet without dropping the identity. TTS 2 is that idea as a flagship. The extra languages (100+ vs Sonic’s 42 native, per each vendor’s docs) matter when the slate is not the 42. Style steering matters when the direction is a sentence (“exhausted, still polite, slightly amused”) rather than a slider.

Credits do not match: 4 and 2 per 1,000 characters. Inworld is half the bill, but do not pick on sticker. Pick on whether the failure mode is a misread code or a drifted persona.

The other Inworld row in the catalog is Inworld TTS 2 Flash, also 2 credits per 1,000 characters: the low-latency tier of TTS-2. This pick is TTS 2 at 2 versus Sonic 3.5 at 4. If you needed the low-latency Inworld tier, that is a different confirm.

Cloning is its own step. On Versely, creating a clone costs no credits (up to 3 saved clones per voice engine), and you pay when the cloned voice speaks, per 1,000 characters at that row’s usual price. Sonic 3.5 already lists voice-clone in its categories at 4 credits per 1,000 characters. If the brief is clone this sample, start from AI voice cloning.

Latency versus character is the split

Sonic when:

The script is operational: numbers, SKUs, URLs, emails, promo codes.
You are iterating many short reads in a session. Fast first-audio is a production benefit even when the viewer never sees it.
The language is inside Cartesia’s 42 native set and you want that native prosody, not a translation-shaped read.
You needed a clone path on the Cartesia stack.
The VO will sit under a silent video plate (Kling Turbo, Hailuo 2.3 audio false) and must be done in volume.

Inworld TTS 2 when:

The voice is a character who has to live for more than one ad: same person, many lines, many moods.
The direction is natural language, not a preset name. “Wry, then suddenly sincere on the last clause.”
The language is outside Sonic’s 42 but inside Inworld’s 100+.
You would throw the take because it stopped being that person at line forty, not because it said “one zero four” wrong.

Neither when:

You needed the mouth to say the line in the generate. Veo 3.1 or Kling O3 Pro, not TTS laid on a closed mouth.
You needed music. Suno is a different card.
You needed captions. TTS is audio. Add captions to video is type.

A legal combo: Hailuo or Turbo silent plate + Inworld or Sonic VO + captions. That is three rows. It is still cheaper than forcing a talking generate to also be the voice talent when the mouth is off-screen.

Do not test them on a pretty paragraph about a sunset. Test Sonic on “Call 1-800-555-0199, code SUMMER-30, visit versely.studio/pricing.” Test Inworld on a 400-word monologue with two emotion turns. Keep the winner of that test. GSC queries are the model names; the click should land on this split, not on a latency essay.

How to run the file like a file

Write the script as the talent should say it. TTS is not a place to paste a shot list. Mark pauses. Spell the brand the way it should be pronounced if it has ever been misread.

On Sonic, leave numbers as numbers unless a test failed. The pitch of 3.5 is that you should not have to phoneticize every SKU. If a heteronym fails, rewrite that one word, not the whole script.

On Inworld TTS 2, put the style in the steering language the card is for. “Same character as the last take, quieter, no smile in the voice.” Do not also paste a cinema prompt. It is not a video model.

Match the video’s mouth policy. If we see teeth, generate speech with the picture or lipsync. If we see hands and a product, a voice file is correct. UGC-shaped ads often want the second: UGC as a talking sound over b-roll, not a generated face. That is TTS.

Article 50: a fully AI voice on a published clip can still be in scope as synthetic audio. Label when you are the deployer. Disclosure template.

The latency deep-dive remains useful if your question is only first-audio milliseconds. If your question is which row to confirm, stay here.

Search is the model name; production is the failure mode

GSC, 7 days to 4 September 2026: “sonic 3.5” at 12 impressions, position 9; “inworld tts-2” at 11 impressions, position 9. People search the SKU. They need the split. A latency essay answers one of those queries and leaves the character job on the table. A character essay leaves the SKU-read job on the table.

Put the names in the title. Put the jobs in the first paragraph. Then test with the script that would actually ship, not with a paragraph about rain. If the ad is a code and a URL, Sonic 3.5. If the ad is a person who has to stay that person, Inworld TTS 2. If the ad is a face saying the line, neither — you are in video.

Gemini 3.1 Flash TTS and ElevenLabs Multilingual sit on the same roster. They are not this pick. Hume Octave 2 ranks in the same GSC export and is not a Versely card today; do not invent it into the catalog. Hume is also sunsetting its TTS API, with access ending on 13 November 2026 (Hume changelog), so it is not a pick for new work.

If you want both voices on one bill, sign up and run the SKU script on Sonic 3.5 and the character script on Inworld TTS 2. Keep the file that survived its own test.

FAQ

Are Sonic 3.5 and Inworld TTS 2 the same price on Versely?

No. Sonic 3.5 is 4 credits per 1,000 characters and Inworld TTS 2 is 2. Sonic 3.5 lists text-to-audio and voice-clone. Inworld TTS 2 lists text-to-audio, 100+ languages, style steering. Pick on the job, not the sticker.

Is this the Sonic latency post?

No. The latency post is about sub-90ms first-audio and who that number is for. This page is Sonic 3.5 versus Inworld TTS 2: technical text and speed versus character and steering. GSC still queries both names.

Can I use these instead of a talking video model?

When the mouth is off-screen, yes: silent plate plus a voice file plus captions. When the mouth is the shot, no. Generate speech with the picture or use lipsync. A cheap VO will not open a closed face.