TL;DR: Public AI music API pricing currently ranges from $0.04 for a 30-second Google Lyria 3 Clip to per-minute pricing for longer compositions. The useful comparison is not just cost per request—it is the cost per accepted, production-ready song after retries, editing, review, and delivery.
AI music APIs are hard to compare because “one generation” does not mean the same thing across providers.
A 30-second clip, a three-minute structured song, and a six-minute audio generation may all be sold as AI-generated music, but they have different billing units, output formats, and production uses.
Current pricing overview
| Provider and model | Public list price | Billing unit | Maximum output | API access |
|---|---|---|---|---|
| Google Lyria 3 Pro | $0.08 | Full song | Up to 3 minutes | First-party API, Preview |
| Google Lyria 3 Clip | $0.04 | 30-second clip | 30 seconds | First-party API, Preview |
Google Lyria 2 (lyria-002) |
$0.06 | 30 seconds | About 30 seconds | First-party API, GA |
| Eleven Music | $0.15 | Generated minute | Up to 10 minutes | First-party API |
| Stable Audio 2.5 | $0.20 | One generation | Up to 3 minutes | First-party API |
| Stable Audio 3.0 | $0.26 | One generation | Up to 6 minutes | First-party API |
| Mureka API | $0.05 | Per song | Model dependent | First-party API |
These prices are not directly interchangeable. Duration, failed-generation billing, retries, and output acceptance rates can change the economics substantially.
How the major APIs bill
Google Lyria
Google currently exposes several pricing routes:
- Lyria 3 Pro: $0.08 per full song, up to three minutes.
- Lyria 3 Clip: $0.04 per 30-second clip.
- Lyria 2: $0.06 per 30 seconds.
The billing unit varies by generation format, as described in Google's Generative AI pricing.
Lyria 3 Pro is aimed at more complete compositions. Lyria 3 Clip is for shorter assets, while Lyria 2 (lyria-002) is the earlier generally available route.
At the maximum advertised Lyria 3 Pro duration, the theoretical cost is:
$0.08 / 3 minutes = $0.027 per minute
That is one of the lowest normalized prices among models with clearly documented maximum durations. It only holds, however, when the full output is useful. A discarded three-minute song still costs $0.08.
Before production deployment, I would verify the current Google Cloud pricing and Lyria documentation, particularly because Lyria 3 routes are listed as Preview.
Eleven Music
Eleven Music is billed at $0.15 per generated minute, according to the ElevenLabs API pricing.
| Duration | Estimated cost |
|---|---|
| 30 seconds | $0.08 |
| 2 minutes | $0.30 |
| 5 minutes | $0.75 |
| 10 minutes | $1.50 |
The API supports longer-form workflows with controls for composition, structure, genre, mood, and instrumentation. Its maximum output is listed as up to 10 minutes.
The per-minute price is higher than the fixed-price alternatives, but that is not automatically a disadvantage. If the output needs fewer retries or less editing, the end-to-end cost can be lower. The Eleven Music API documentation is the relevant reference for longer-form generation behavior.
Stable Audio
Stability AI uses credits, with one credit equal to $0.01.
The public pricing is:
- Stable Audio 2.5: 20 credits, or $0.20 per generation, for up to three minutes.
- Stable Audio 3.0: 26 credits, or $0.26 per generation, for up to six minutes.
That produces these theoretical floors:
| Model | Maximum-duration calculation | Theoretical price floor |
|---|---|---|
| Stable Audio 2.5 | $0.20 / 3 minutes | $0.067/min |
| Stable Audio 3.0 | $0.26 / 6 minutes | $0.043/min |
The important word is theoretical. A six-minute generation is only inexpensive if most of those six minutes survive review. If half the output is unusable, the effective cost per usable minute is much higher.
Mureka
Mureka's API pricing page currently lists pricing starting at $0.045 per song. The comparison table above uses the commonly presented $0.05 per-song headline price.
That number is difficult to normalize against minute-based and maximum-duration pricing. Before using it for a real budget, I would confirm:
- The exact model.
- Output duration.
- Generation workflow.
- Concurrency limits.
- Licensing terms.
A per-song price is not necessarily cheaper than a three-minute or six-minute generation until the output contracts are equivalent.
Suno and developer API pricing
Suno is often included in AI music comparisons, but its public pricing page focuses on creator subscriptions, credits, model access, and commercial-use features. It does not currently list a first-party per-call developer API price comparable to the APIs above.
I therefore exclude Suno from the API pricing table. A creator subscription should not be compared directly with documented production API billing.
Normalizing the advertised prices
Among models with clearly documented maximum durations, the theoretical per-minute ranking looks like this:
| Route | Maximum-duration calculation | Theoretical price floor |
|---|---|---|
| Lyria 3 Pro | $0.08 / 3 minutes | $0.027/min |
| Stable Audio 3.0 | $0.26 / 6 minutes | $0.043/min |
| Stable Audio 2.5 | $0.20 / 3 minutes | $0.067/min |
| Lyria 3 Clip | $0.04 / 0.5 minute | $0.080/min |
| Lyria 2 | $0.06 / 0.5 minute | $0.120/min |
| Eleven Music | Direct per-minute billing | $0.150/min |
This is a budgeting comparison, not a quality ranking.
A collection of independent 30-second clips is not equivalent to one coherent full-length song. Likewise, a six-minute generation only has a low effective rate when the output is usable throughout.
Mureka is intentionally absent from this table because its published per-song price cannot be normalized without confirming the model and output duration.
The metric that matters: cost per accepted output
For production, I care less about cost per API call than cost per song that passes review.
The broader calculation is:
cost per usable song =
(generation spend
+ retry spend
+ editing spend
+ storage and delivery cost
+ human review cost)
/ accepted songs
For a generation-only estimate:
expected generation cost per accepted song =
generation cost per attempt / acceptance rate
Consider a two-minute Eleven Music generation:
| Item | Calculation | Cost |
|---|---|---|
| One two-minute generation | 2 × $0.15 | $0.30 |
| Expected attempts at 25% acceptance | 1 / 0.25 | 4 attempts |
| Generation cost per accepted song | 4 × $0.30 | $1.20 |
| Example post-processing and storage | Planning assumption | $0.06 |
| Estimated workflow cost | $1.20 + $0.06 | $1.26 |
The 25% acceptance rate and $0.06 post-processing allowance are illustrative assumptions, not ElevenLabs benchmarks.
Using that same hypothetical 25% acceptance rate, a Stable Audio 3.0 generation costs:
$0.26 / 0.25 = $1.04 per accepted generation
That does not prove Stable Audio is cheaper. The comparison only works when the outputs satisfy the same requirements for duration, structure, licensing, prompt adherence, and quality.
A cheaper API call can become more expensive if it produces more rejected generations or requires more manual cleanup. I would track both:
- Cost per generation.
- Cost per accepted output.
The second number is the one I would use for production planning.
A production comparison checklist
Use equivalent briefs
Run the same real workloads through every provider rather than comparing hand-picked demos. Useful test cases include:
- A 30-second background track for social video.
- A two-minute structured song.
- A loopable soundtrack for a game or application.
- A brand-safe advertising track with specified instruments, tempo, and exclusions.
Measure the whole asynchronous workflow
Request latency is not completion latency. Record:
- Queue time.
- Generation time.
- Callback or webhook delay.
- Download time.
- p50 completion time.
- p95 completion time.
- Provider errors and timeouts.
Capture billing and editorial data
For each request, store:
- Provider and model version.
- Requested and returned duration.
- Number of outputs.
- Amount billed.
- End-to-end completion time.
- Provider success or failure.
- Editorial acceptance or rejection.
- Retry reason.
- Human review and editing time.
This data is enough to calculate cost per accepted output instead of relying only on public pricing pages.
Verify licensing before scaling
Commercial-use terms can differ by provider, model, plan, and distribution channel. Check coverage for:
- Advertising.
- Film and television.
- Games.
- Podcasts.
- Streaming distribution.
- Client work.
- High-volume automated generation.
Also check attribution requirements, artist-imitation restrictions, output ownership, and enterprise licensing.
Test operational edge cases
Before committing to a provider, confirm:
- Whether failed jobs are billed.
- How moderation failures are handled.
- Webhook retry behavior.
- Idempotency support.
- Output URL expiration.
- Model version pinning.
- Concurrency and rate limits.
- Fallback behavior.
A small per-generation saving can disappear if the provider creates additional operational work.
Matching APIs to workloads
| Workload | First route to test | Why | Main check |
|---|---|---|---|
| Short instrumental assets | Lyria 2 | Established short-form API | 30-second limitation |
| Short music clips with newer capabilities | Lyria 3 Clip | Low per-clip price | Preview availability |
| Structured full songs | Lyria 3 Pro | Complete compositions up to three minutes | Preview status and acceptance rate |
| Long-form audio assets | Stable Audio 3.0 | Up to six minutes at fixed generation pricing | Useful output across full duration |
| Detailed structured compositions | Eleven Music | Composition and section controls | Licensing and acceptance rate |
| Low-cost song-generation experiments | Mureka API | Low published per-song price | Exact model and output contract |
I would not route every workload to one model. Short-form assets, loopable audio, structured songs, and longer compositions can have very different quality and cost profiles.
A unified multi-model API such as CometAPI can be useful when the application needs to test or route among several providers, but the provider's own pricing, model availability, and terms still need to be verified.
A 50-prompt benchmark
Once the shortlist is small, I would run a controlled 50-prompt evaluation using actual production briefs:
- 15 short background tracks.
- 10 loopable game or application tracks.
- 10 structured songs.
- 10 brand-constrained advertising tracks.
- 5 difficult multilingual or reference-based tasks.
For each provider, calculate:
acceptance rate =
accepted outputs / completed outputs
cost per accepted output =
total API spend / accepted outputs
p95 delivery time =
95th percentile of end-to-end completion time
I would keep the workload categories separate rather than collapsing them into one score. An API that performs well for short background music may be a poor choice for longer structured compositions.
The goal is to find the best route for each workload, not to declare a universal winner.
FAQ
How much does an AI music API cost in 2026?
Current public prices include $0.04 for a 30-second Google Lyria 3 Clip, $0.08 for a Lyria 3 Pro song up to three minutes, $0.15 per generated minute for Eleven Music, and $0.20 to $0.26 per Stable Audio generation. Mureka's public API pricing starts at $0.045 per song.
Which API is cheapest?
There is no universal answer because the billing units and output formats differ. Among models with clearly documented maximum durations, Lyria 3 Pro has a theoretical floor of approximately $0.027 per minute when the full three-minute output is used. Mureka also publishes a low per-song price, but its exact output contract should be verified first.
Does Suno have a developer API?
Suno's public pricing currently describes creator subscriptions, credits, and model access rather than a first-party per-call developer API price comparable to the APIs in this comparison. That is why it is excluded from the main table.
How should I calculate real generation cost?
Add generation, retry, editing, storage, delivery, and human review costs, then divide by the number of outputs that pass your acceptance criteria. A low acceptance rate can make a nominally cheap API expensive in production.
What should I test before choosing a provider?
Use identical prompts and acceptance criteria to compare prompt adherence, audio quality, song structure, duration, completion time, acceptance rate, retry behavior, rate limits, model stability, commercial-use terms, and cost per accepted output.
Originally published at cometapi.com
Top comments (0)