HomePricingStronger than Kugel 3.
A quarter cheaper.
Our new text-to-speech model gets short replies, IPA pronunciation, pauses and spelled-out codes right far more often, at 75% of the price of Kugel 3.
The small things
callers notice first.
A voice agent is judged on its details: a one-word answer, a name said correctly, a pause that lets the caller think, an order number read out character by character.
Kugel 3.5 was trained on exactly these. It keeps the voices, languages and API of Kugel 3, and gets the details right far more often.
“Yes.” “Okay.” “One moment.” A single word gives a speech model almost nothing to go on, so this is where models slip most. Kugel 3.5 says them cleanly, in the same voice every time.
95.5%of short German replies said correctly, up from 84.8% with Kugel 3
Give a name, a brand or a place its IPA transcription in a pronunciation dictionary, and Kugel 3.5 says it that way instead of guessing from the spelling.
65.2%of German IPA pronunciations hit, up from 36.7% with Kugel 3
{
"word": "Nguyen",
"ipa": "ŋwiən"
}A break tag sets the pause before the next sentence. Kugel 3.5 stays close to the length you ask for. ElevenLabs v4 Turbo has no timed pause and pauses about the same whatever you write.
31.8%of pauses within 50 ms of the requested length, up from 18.6% with Kugel 3
<break time="200ms"/> | |
|---|---|
| Kugel 3.5 | 197 ms |
| ElevenLabs v4 Turbo | 555 ms |
| Cartesia Sonic 3.6 | 186 ms |
<break time="400ms"/> | |
| Kugel 3.5 | 370 ms |
| ElevenLabs v4 Turbo | 566 ms |
| Cartesia Sonic 3.6 | 401 ms |
<break time="500ms"/> | |
| Kugel 3.5 | 456 ms |
| ElevenLabs v4 Turbo | 563 ms |
| Cartesia Sonic 3.6 | 495 ms |
Order numbers, licence plates and booking codes inside a spell tag are read out one character at a time, in the order you wrote them.
99.4%of spelled-out codes read correctly, up from 95.4% with Kugel 3
<spell>KX7Q2M</spell>| Test | Kugel 3.5 | ElevenLabs v4 Turbo | Cartesia Sonic 3.6 | Deepgram Aura-2 | Qwen3-TTS 1.7B | Kugel 3 |
|---|---|---|---|---|---|---|
| Short replies, voice 1887 for Kugel 3.5, Kugel 3 and Qwen3-TTS | ||||||
| Words pronounced with the wrong accent given the reference voicelower is better | 0% (best) | 4.3% | 0.5% | 12.8% | 15.7% | 0.7% |
| Short words said wronglower is better | 3% (best) | 7.7% | 6.1% | 20.6% | 10.6% | 16.8% |
| Abrupt endingslower is better | 0.7% | 0% (best) | 3.2% | 42.4% | 75% | 10.9% |
| Delay before the first wordlower is better | 40 ms (best) | 60 ms | 160 ms | 340 ms | 320 ms | 100 ms |
| Following your markup | ||||||
| 200 ms pauses within 50 mshigher is better | 46.3% (best) | 1.8% | 36% | 15.5% | 20.5% | 24.8% |
| German IPA pronunciationhigher is better | 65.2% (best) | 57.3% | 0.8% | 29.4% | 10.2% | 36.7% |
| English IPA pronunciationhigher is better | 75% (best) | 65.6% | 6.2% | 32.5% | 5.4% | 53.3% |
| Reading content right | ||||||
| Numbers read exactlyhigher is better | 97.1% (best) | 95.4% | 94% | 36.1% | 18.4% | 62.9% |
| Numbers and codes read in their groupshigher is better | 90.8% (best) | 71.9% | 82.6% | 0.9% | 1.5% | 20.6% |
KugelBench, Kugel 3.5 on 08.10.2026, the other models between 29.09.2026 and 06.10.2026. Every model spoke the same texts and was scored by the same automatic judges. Qwen3-TTS 1.7B is the open base model with our voices cloned from a short sample. German IPA is scored by the xlsr judge, English IPA by the ZIPA judge. Numbers: 1,120 paragraphs, correct only if every digit is. Short replies: 30 German words, 4,500 clips of voice 1887 for Kugel 3.5, Kugel 3 and Qwen3-TTS, 9,000 clips across two German voices of each other model. The delay is the 90th percentile. A pause counts as correct within 50 ms of the requested length; ElevenLabs v4 Turbo has no timed pause, so its pause tag was used.
€0.07per minute of audio
excl. VAT · €0.0833 incl. 19% VAT
Kugel 3 stays available at its current price.
€0.0525per minute of audio
excl. VAT · €0.062475 incl. 19% VAT
The voices, languages and API you already use, billed per second of audio.
Try Kugel 3.5Pay-as-you-go prices. Consumers in other EU countries pay their local VAT rate, shown before payment. Enterprise plans are priced separately.
Kugel 3.5 is our newest text-to-speech model. It builds on Kugel 3 and was trained to get the details of phone conversations right: short replies, IPA pronunciation, timed pauses and spelled-out codes.
On our benchmark of short German replies, Kugel 3.5 says 95.5% correctly, against 84.8% for Kugel 3. It also follows IPA transcriptions, requested pause lengths and spelled-out codes more reliably. The table above compares it with ElevenLabs v4 Turbo and Cartesia Sonic 3.6.
No. Set the model to kugel-3.5 in your request. Voices, languages, dictionaries and the break and spell tags work the same way as with Kugel 3.
€0.0525 per minute of audio, excl. VAT, billed per second. That is 25% less than Kugel 3 at €0.07.
Kugel 3.5 is available to every account. In the API, set model_id to kugel-3.5; in the dashboard, pick Kugel 3.5 from the model list.