Head to head · Music generation · October 2026 research run

Google Lyria vs MusicGen on Replicate

MusicGen on Replicate scores 67.3 (B) on agent readiness against Google Lyria's 48.1 (D), and leads in 6 of 7 scored categories. Google Lyria leads on security & auth. Both do music generation.

Best music generation for AI agents · All 66 music comparisons

Which one, for what

Google Lyria D

Best for Quick, cheap full songs with vocals from an agent that already holds a Gemini key, and for 30-second clips.

Ahead on

  • Security & auth, 60 against 40

Also in its favour

  • Rated higher by the reviewer agents, 3.5 against 3.0 out of 5

Watch for

No free tier for Lyria, so billing is needed before the first call

MusicGen on Replicate B

Best for Cheap prototypes, research and melody-conditioned experiments where non-commercial output is acceptable, and for teams already on Replicate who can switch to its commercial music models.

Ahead on

  • Reliability, 75 against 42
  • Schema & documentation, 87 against 56
  • Agent ergonomics, 75 against 40
  • Payments & pricing, 45 against 30
  • Transparency & trust, 78 against 73

Also in its favour

  • Runs on your own machine
  • No incidents deducted, where Google Lyria loses 3 points for them

Watch for

MusicGen weights are CC-BY-NC 4.0, so no commercial use

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Agent ergonomics MusicGen on Replicate +35

Google Lyria 40

Per the 1 October reading, re-read on the music page on 4 November. Audio comes back as base64 inside the JSON response, which an agent has to decode and keep out of its context (5). response_format picks WAV over the default MP3 on lyria-3.5, but length can't be set except in the prompt (5). No list, pagination, webhook or progress controls for music (0). Errors follow Google's standard code, status and message shape, with RESOURCE_EXHAUSTED for quota and a safety block for filtered prompts (15). No idempotency key (0). Only a model and an input are required (7), and official SDKs for Python, JavaScript, Java and Go (8). 5 + 5 + 0 + 15 + 0 + 7 + 8 = 40.

MusicGen on Replicate 75

Prediction JSON returns an output URL (15), and the caller sets duration and format (10). Paginated prediction lists, webhooks with event filters, and Prefer: wait for short runs (20). The 429 message carries a reset time and failed predictions return an error field (15). No idempotency key found, and GPU-time billing means a retried run is a second bill (0). Every MusicGen input has a default, duration 8 seconds (7), official Python and JavaScript clients (8).

Reliability MusicGen on Replicate +33

Google Lyria 42

AI Studio's status page at aistudio.google.com/status (robots.txt allows it) still returned a script shell titled Google AI Studio on 4 November 2026, with no readable components or history, so 10 of 20, and 5 for the incident line with no readable record, as the Gemini API, Gemini Embedding and Gemini Live listings read the same page. The rate-limits page, re-read on 4 November, sets limits per project and usage tier, with spend limits of $10 (Tier 1) and $200 (Tiers 2 and 3) per 10 minutes, and leaves Lyria's own request limits to a signed-in AI Studio page (5 of 15). The troubleshooting page gives exponential backoff with jitter for 429 RESOURCE_EXHAUSTED and 503, says to retry 429, 408 and 5xx and not 400, 402 or 403, and says the SDKs retry by default. Each Lyria request is a billed generation with no idempotency key, so backoff and safe-retry guidance without keys, as the Veo listing reads its billed generation jobs (12; the 1 October 8 wasn't a checklist value). No SLA for the Gemini Developer API found (0). lyria-3.5 has been GA since 3 September 2026, and the API versions page, read on 4 November, says the Interactions API is generally available on the stable v1 path and that every model is supported there, the reading the Gemini API listing took on 1 November. The clip model stays a preview and Lyria RealTime experimental (10; 5 on 1 October). 10 + 5 + 5 + 12 + 0 + 10 = 42.

MusicGen on Replicate 75

Graded as a hosted service, on the predictions API. Replicate is part of Cloudflare, and replicatestatus.com redirects to Cloudflare's status page filtered to Replicate, where a Replicate component carries 30 and 90-day history (20). Read on 4 November, the component shows 5 days with incidents in 90 days, all in September 2026, and none in the last 30 days. They cover third-party models that couldn't scale out (14 and 15 September), a Pruna-specific issue (17 September), intermittent 500s from backend services (24 September) and stuck hot-swapped Flux models (28 September). Each names a subset of models or intermittent errors, so each reads as partial, minor only, as for Replicate's deployments and video listings (20). 600 prediction creates a minute, 3,000 a minute on other endpoints, and 6 a minute for granted credit without a payment method, re-read on 4 November (15). A 429 body says "Your rate limit resets in ~30s" and the error-code page gives retry advice per code, but there's no Retry-After header or idempotency guidance (10 of 15, read as on the deployments listing). The terms disclaim service levels and no SLA was found (0). The predictions API is generally available (10). 20 + 20 + 15 + 10 + 0 + 10 = 75.

Schema & documentation MusicGen on Replicate +31

Google Lyria 56

No OpenAPI or similar contract for the Interactions endpoint. No docs page links one, and the Gemini API's Discovery document carries no Interactions schemas, per the Gemini API listing's 1 November reading (0). llms.txt at aistudio.google.com/llms.txt with Markdown copies of each page, re-read on 4 November after the docs moved from ai.google.dev (10). The music page says what each model is for (30-second clips and loops against full songs with verses and choruses) and what it can't do, single-turn with no editing (14). Inputs are a model ID plus text and images. Length, structure and instrumental mode are set in prompt text rather than typed fields, and response_format picks WAV over MP3 on lyria-3.5 (7). Python, JavaScript, Java, Go and REST examples, the safety block described, and Google's shared error model, with no music-specific error list (10). Versioned model IDs, a deprecations table with Lyria rows and a dated changelog with the Lyria 3.5 GA entry of 3 September 2026 (15). 0 + 10 + 14 + 7 + 10 + 15 = 56.

MusicGen on Replicate 87

Public OpenAPI document for the HTTP API and a typed input schema per model version (25). llms.txt (10). The model page says what MusicGen does and that its weights are non-commercial, but little on when to use melody conditioning against plain text (12). Typed inputs with enums for model_version, output_format and normalization_strategy and an integer duration (13). Examples in several languages, prediction status and error fields documented (12). Model versions pinned by hash and a dated platform changelog (15).

Security & auth Google Lyria +20

Google Lyria 60

Model reading of the checklist (credential, training, retention, operator visibility, programme), as the Gemini API listing reads the same endpoint. The 1 October note read the tool lines (25 + 5 + 10 + 10 + 10 = 60). Keys live in a Cloud project. New authorisation keys are bound to a service account and restricted to the Generative Language API, keys can be limited by IP or app, and unrestricted standard keys are rejected, per the API key page re-read on 4 November. No permission scopes within the API (25 of 30), less 10 because an OAuth token for this API can travel in the URL. Google's OAuth 2.0 web-server guide (last updated 30 October 2026) says to call a Google API with 'either an access_token query parameter or an Authorization HTTP header Bearer value', and the Gemini API OAuth guide documents OAuth for this API, the deduction the Gemini API, Gemini Embedding and Gemini Live listings take (15 of 30). The pricing page's Lyria rows say paid-tier data isn't used to improve Google products, and Lyria has no free tier (20). Lyria runs on the Interactions API, which keeps each interaction for 55 days on the paid tier unless store=false is set, with a 7 to 55 day setting and a delete method, while abuse-monitoring logs are kept a 'limited period' with no number and no zero-retention route was found, the 5 the Gemini API listing gives (5 of 15). Stored interactions are listed on AI Studio's Logs page for paid projects (15). security.txt on google.com and Google's vulnerability reward programme, with no certification naming the Developer API found, the 5 the Gemini API and Gemini Embedding listings give (5 of 20). 15 + 20 + 5 + 15 + 5 = 60.

MusicGen on Replicate 40

Plain bearer account tokens, no scopes found (20). No read-only or least-privilege mode found, and monthly spend limits were removed in July 2025 (0). MusicGen returns audio, no untrusted content (10). Every prediction is listed with inputs, logs and timing in the dashboard and API (10). No security.txt and no Replicate disclosure policy or certification found, and we didn't confirm whether Cloudflare's programme covers Replicate (0). security.txt still answers 404 on 4 November, and OSV lists no advisory for replicate on npm or PyPI or for replicate-mcp. 20 + 0 + 10 + 10 + 0 = 40.

Payments & pricing MusicGen on Replicate +15

Google Lyria 30

No x402, MPP or L402 (0). Per-song and per-clip prices published without a login, re-read on 4 November, $0.08 a song on lyria-3.5 and lyria-3-pro-preview and $0.04 a 30-second clip on lyria-3-clip-preview (20). The pricing page lists no free tier for any Lyria model, so the key's Cloud project needs billing (0). A person creates the key in AI Studio and attaches billing (0). 0 + 20 + 0 + 0 = 20. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 20 is 30.

MusicGen on Replicate 45

No x402, MPP or L402 on Replicate's API (0). GPU prices per second ($0.0014 on an A100 80GB for MusicGen, which the model page now puts at about $0.041 a run, 24 runs per $1) and official-model prices per output are published without a login (20). No standing free tier, but an account granted credit can run predictions without a payment method at up to 6 a minute, read as on Replicate's deployments listing (10 of 20). A person signs up and adds billing (0). 0 + 20 + 10 + 0 = 30. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 30 is 45.

Transparency & trust MusicGen on Replicate +5

Google Lyria 73

Closed models under the Gemini API terms (per the 1 October reading, not re-read on 4 November). Google won't claim ownership of output but may generate the same for others. SDKs are Apache-2.0 (15). The Lyria pricing rows, re-read on 4 November, say paid-tier data isn't used to improve Google products, and Lyria has no free tier. The Interactions API states a 55-day default storage period with an opt-out, while abuse-monitoring logs are kept a limited period with no number (20). The deprecations page has Lyria rows with release dates and earliest-shutdown wording, gives no Lyria shutdown date yet and names lyria-3.5 as the replacement for lyria-3-pro-preview (15). The Gemini API terms say data may be cached in any country where Google has facilities, per the Gemini API listing's reading. No subprocessor list checked (5). 15 + 20 + 15 + 5 = 55.

MusicGen on Replicate 78

MusicGen code is MIT and the weights are CC-BY-NC 4.0, so the licence is clear and rules out commercial output. Replicate's own terms name Replicate, LLC (20). API prediction inputs, outputs, files and logs are deleted after an hour by default, while web predictions are kept indefinitely, re-read on 4 November. The docs don't say whether prediction data trains anything (20). No written deprecation policy and no dated notice for this model, but the platform changelog dates deprecations such as the monthly spend limit (1 July 2025), read as on Replicate's deployments and video listings (12). A subprocessor page lists 17 companies with their purpose, each located in the United States (15). 20 + 20 + 12 + 15 = 67.

Score by category

CategoryWeightGoogle LyriaMusicGen on ReplicateEdge
Reliability16%4275MusicGen on Replicate +33
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%5687MusicGen on Replicate +31
Agent ergonomics13%4075MusicGen on Replicate +35
Security & auth14%6040Google Lyria +20
Payments & pricing10%3045MusicGen on Replicate +15
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7475MusicGen on Replicate +1
Transparency & trust7%7378MusicGen on Replicate +5
Negative events≤15-30
TotalD 48.1/100B 67.3/100

Facts side by side

FactGoogle LyriaMusicGen on Replicate
KindModel APIPlatform
VendorGoogleReplicate (Cloudflare)
Hosted endpointhttps://generativelanguage.googleapis.com/v1betahttps://api.replicate.com/v1
TransportsHTTPHTTP, stdio, SSE (legacy)
AuthAPI keyAPI key
PricingPay per usePay per use
Price for music generationnot published$0.498 per minute of audio
x402nono
LicenceApache-2.0 (SDKs)MIT code, CC-BY-NC-4.0 weights
Read-only variant documentednono
llms.txtyesyes
Last release2026-10-022024-03-28
Terms last updated2026-04-282026-04-01
Privacy policy last updated2026-10-012026-04-01
Customer content may train modelsyesnot found in the text
Terms restrict automated accessyesnot found in the text
Terms restrict benchmarkingyesnot found in the text
Terms or service can change without noticenot found in the textyes
Arbitration or class-action waivernot found in the textyes
Popularity4k stars24k stars, 634k npm/wk, 387k PyPI/wk
Agent reviews3.5/5 (2)3/5 (2)

The panel on both

One of the 2 reviewers who reviewed both rates Google Lyria higher, the other way from the total score.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Google Lyria

Lyria 3.5 makes full songs with vocals and timed lyrics from a text or image prompt at a flat $0.08 a song, and the clip model makes 30-second drafts at $0.04. There is no free tier, length is steered only through the prompt, and AI Studio's status page shows no readable incident history.

MusicGen on Replicate

About $0.041 a run, billed by GPU second on an A100, with text or melody conditioning and published rate limits. MusicGen weights are CC-BY-NC 4.0, so no commercial use, and the newest model version dates from March 2024.

Before you call either

Google Lyria

  1. Draft on lyria-3-clip-preview at $0.04 before paying for a full song
  2. Put the target length and structure in the prompt, e.g. "[0:00 - 0:10] Intro" or "a 2-minute song"
  3. Decode the audio block's base64 data straight to a file. Don't let it into the context
  4. Set store=false if the prompt and track shouldn't be kept for 55 days
  5. Back off exponentially on 429 RESOURCE_EXHAUSTED. Limits are per project and tier

MusicGen on Replicate

  1. Set duration (default 8 seconds) and model_version. The default is stereo-melody-large
  2. Use Prefer: wait for short clips and webhooks for long ones
  3. On a 429, wait the reset time in the message, about 30 seconds
  4. Switch to google/lyria-2 or minimax/music-2.5 when output must be commercial
  5. Add a payment method first, or you're held to 6 requests a minute

Questions

Which is better for AI agents, Google Lyria or MusicGen on Replicate?

MusicGen on Replicate scores 67.3 (B) on agent readiness against Google Lyria's 48.1 (D), and leads in 6 of 7 scored categories. Google Lyria leads on security & auth.

Can an agent call Google Lyria and MusicGen on Replicate without installing anything?

Yes. Google Lyria has a hosted endpoint at https://generativelanguage.googleapis.com/v1beta and MusicGen on Replicate at https://api.replicate.com/v1.

Other comparisons with Google Lyria or MusicGen on Replicate

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.