이미지: X — 미디어·생성AI
Summary
- Pika announced four audio models — Pika Soundtrack, Music, SFX, and Speech — and said they will be available only through Pika API Club
- Pika claimed the models are cheaper than every audio model on the market, up to 20 times cheaper, with the only disclosed figure being Soundtrack's $0.617 per second
- Pika named Hunyuan Foley as the comparable model for turning video into sound and said its own Soundtrack is twice as cost-efficient
- 발표
- Pika, 2026년 8월 14일 X 게시
- 모델 수
- 4개 — Pika Soundtrack, Pika Music, Pika SFX, Pika Speech
- 제공 경로
- Pika API Club 전용
- 가격 주장
- 시장의 모든 오디오 모델보다 저렴, 최대 20배
- Soundtrack 단가
- 초당 0.617 (원문 표기 0.617 / seconds)
- 비교 모델
- Hunyuan Foley 대비 비용 효율 2배, Pika는 이 모델을 비교 가능한 유일한 video-to-audio 모델로 지목
- 저가 근거
- Pika는 학습·추론 효율 개선을 이유로 들었다
- 예고
- Pika는 생성 미디어 접근성을 높이는 후속 발표를 예고했다
Adding sound to video has long been a job held by human hands. Foley work — matching the timing of a single footstep or a door closing to what's on screen — requires precision down to the second. The four audio models Pika unveiled on August 14 propose turning that entire process into an API call. And what Pika led with in the announcement wasn't performance, but price.
Sound split into four
Pika said it released four frontier foundation models covering the full range of generative sound. The division of labor is clear from the names alone. Soundtrack handles sound matched to video, Music handles music, SFX handles sound effects, and Speech handles voice. Rather than combining everything into one all-purpose model, the lineup is split by purpose.
| Model | Domain | Disclosed figures |
|---|---|---|
| Pika Soundtrack | Video-to-audio | $0.617/second · 2x cost-efficiency vs. Hunyuan Foley |
| Pika Music | Music generation | — |
| Pika SFX | Sound effect generation | — |
| Pika Speech | Speech generation | — |
Of the four models, only Soundtrack came with a specific figure. Pika listed 0.617 per second, though the post did not specify a currency unit. Pika named Hunyuan Foley as the only comparable model for generating sound from video input, and claimed its own Soundtrack is twice as cost-efficient. Hunyuan is a model family built by Tencent.
An announcement led by price
Pika said it built the models to be cheaper than every audio model on the market, up to 20 times cheaper. Such statements are usually followed by a footnote listing conditions. Instead, Pika added an asterisk and wrote "no disclaimers, seriously" (Pika, X). It's a half-joking line, but it also amounts to a declaration that the numbers come with no caveats attached.
Pika attributed the lower pricing to improvements in the team's training and inference efficiency. Inference — the stage where an already-trained model actually produces output — translates directly into unit cost for modalities like audio that are billed by length.
How to get started
Where to begin. There's a single entry point: Pika API Club. Pika stated plainly that the four models are available only there. An API is a channel through which programs call on one another, so "released via API" means other apps and editing tools can now plug into these models. In other words, this launch isn't aimed at a general user clicking a button on the web — it targets developers and production pipelines first.
Steps.
- Sign up on Pika API Club to get API access. The four audio models appear together in the list of available models.
- Choose a model based on the sound you want to create. Use Soundtrack for sound matched to video, Music for background music, SFX for individual sound effects, and Speech for dialogue or narration.
- Soundtrack takes video as input and generates sound matched to that scene. Billing is based on output length in seconds, so it's safer to test with short clips first.
- Layer the resulting audio track onto an editing timeline and re-call the model only for sections that don't sync with the picture.
What you can try. For example, feeding a 30-second self-shot product video into Soundtrack produces the sound of hands touching a box along with ambient room noise in a single pass. For a short ad, you could pull a background track from Music and narration from Speech separately, then layer them together. For work needing hundreds of sound effects, like a game prototype, calling SFX repeatedly to fill out a list fits the per-second pricing structure.
Why price comes first
It's not unusual this year for pricing to headline a generative AI announcement. Grok 4.6, released on August 12, priced at $2 input / $6 output per million tokens, undercutting both Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30). The same day, DeepSeek rolled out official pricing for DeepSeek-V4-Pro-0813 at $0.435 input / $0.87 output per million tokens on a cache miss. This is the same trend that played out in language models now arriving in audio.
Audio is billed differently from language models — by the second, not the token. That means for work where volume adds up, like a five-minute video or the hundreds of sound effects a game requires, differences in unit price directly change what's feasible to produce.
What actually changes
What changes immediately is for teams building their own video production pipelines. Teams that had already automated video generation but were still bolting on sound through separate tools and separate bills can now call soundtrack, music, sound effects, and speech together through a single API channel. Developers building editing tools or video services now have another option for adding sound generation to their products.
Pika said this audio model lineup is just one of several things it's been preparing behind the scenes, and promised to keep releasing ways to make generative media accessible to more people. The real quality of the four models will ultimately be judged by running different models side by side on the same footage. Separate from per-second pricing, how many frames the sound drifts from the picture is the kind of problem that catches the eye in foley work before price ever does.
