What Is Muse Spark, and What Is the Meta Model API?
Muse Spark is Meta's foundation model, released on 8 April 2026 by Meta Superintelligence Labs under chief AI officer Alexandr Wang. The Meta Model API, in public preview since 9 July 2026, is the self-serve endpoint that sells access to it, and it is compatible with the OpenAI SDK.
That last detail is the one to read twice. OpenAI SDK compatibility means adopting Muse Spark is a base URL swap and a new key, not a rewrite, and it means leaving is just as cheap. Low switching cost is the best argument for testing it on a real workload.
Meta shipped the family in fast succession across 2026:
| Date (2026) | Release | What it is |
|---|---|---|
| 8 April | Muse Spark | The foundation model |
| 7 July | Muse Image | Image generation |
| 9 July | Muse Spark 1.1 + Meta Model API | Self-serve public preview |
| July | Muse Video | Previewed |
| 5 August | Muse Code | Coding agent for large codebases |
| 8 September | Muse | Consumer personal AI agent |
Five months, six releases. For a founder only two questions matter: what it costs to run, and whether anyone subsidises the bill. AI Perks tracks $7.7M in credits across 194 companies, and Meta is not one of them.
Meta Model API Pricing in Full
Muse Spark 1.1, 1.2 and 1.3 all price at $1.25 per million input tokens and $4.25 per million output tokens. Images are $0.01 each, and voice transcription is $3.00 per 1,000 minutes.
These are Meta's own published developer prices, not estimates:
| Product | Price |
|---|---|
| Muse Spark 1.1 / 1.2 / 1.3 (standard) | $1.25 per 1M input, $4.25 per 1M output |
| Muse Spark (Contributor tier) | $0.10 per 1M input, $0.20 per 1M output |
| Muse Image | $0.01 per image |
| Muse Voice Transcribe | $3.00 per 1,000 minutes |
Note that the three model versions share one price. Meta has not published a premium rate for the newer revisions, so on the published card, moving from 1.1 to 1.3 costs nothing extra.
Translated into a budget, $100 of standard-tier spend buys roughly 80 million input tokens, or 23.5 million output tokens, or 10,000 images, or about 555 hours of transcription. A single call carrying 10,000 input tokens and returning 2,000 output tokens costs about 2.1 cents. Run that a thousand times a day and you are at roughly $630 a month before anything else.
That is exactly the size of bill that startup credit programs are designed to erase.
The Contributor Tier: 12x Cheaper on Input, 21x on Output
Meta's Contributor tier charges $0.10 per million input tokens and $0.20 per million output tokens, which is 12.5 times cheaper on input and 21 times cheaper on output. The price of that discount is that Meta trains on your prompts and completions.
The same 10,000-in, 2,000-out call that costs 2.1 cents on the standard tier costs about 0.14 cents on Contributor. A workload running $630 a month on standard rates lands near $42 on Contributor. For a pre-revenue team burning runway, that gap is not a rounding error, it is the difference between shipping and not shipping.
Whether it is a good trade depends entirely on what is in your prompts:
Reasonable candidates. Public data enrichment, content generation from your own marketing copy, classification over non-sensitive text, internal prototypes, and anything you would be comfortable seeing summarised in a model's future output.
Bad candidates. Anything carrying customer PII, health or financial records, proprietary source code, unreleased product plans, or data covered by a customer contract that restricts subprocessing. Your enterprise customers' data protection addendum almost certainly makes this choice for you.
The honest framing is that Contributor is a data-for-discount barter, not a startup program. A real credit program leaves your data alone and pays the bill anyway, which is what the programs listed at AI Perks do.
Muse the Agent Is Not Muse Spark the Model
Muse, launched 8 September 2026, is Meta's consumer personal AI agent. Muse Spark is the underlying foundation model sold through the API. They are separate products with separate pricing, and conflating them is the most common error in coverage of both.
The consumer agent proactively completes tasks rather than answering questions: booking, filling forms, sending emails, making purchases. It runs on a Meta-cloud virtual machine with a visible browser, so you watch it work rather than trusting a log. Its permissions layer, which governs what the agent is allowed to do on your behalf, is called Sentinel.
It is available on iOS, Android, WhatsApp and on the web at muse.ai, with AI glasses listed as coming soon.
| Muse consumer plan | Price |
|---|---|
| Free | $0 |
| Power | $20 per month |
| Maximum | $100 per month |
Meta differentiates these tiers by usage volume rather than by feature access, which is a meaningfully different bet from the competition. It means the free tier is the full product with a ceiling rather than a crippled version. Meta has not published where that ceiling sits, and any article quoting a specific free-tier task limit is guessing.
None of this pricing applies to your application. If you are building software, you pay the Meta Model API rates above, not $20 a month. The consumer subscription buys you a personal agent, not developer quota.
The muse.ai Name Collision, Cleared Up
The muse.ai domain now serves Meta's consumer agent, but it previously belonged to an independent video hosting and search company that rebranded to Skiv on 3 March 2026. Microsoft's Muse is a third, entirely unrelated thing.
Three separate products, one name, and search results that mix all three:
| Name | Who | What it actually is |
|---|---|---|
| Muse / Muse Spark | Meta Superintelligence Labs | Consumer agent plus foundation model, 2026 |
| Skiv (formerly muse.ai) | Independent, founded by Antonio Roldao | Video hosting and search, rebranded 3 March 2026 |
| Muse (WHAM) | Microsoft, with Xbox Game Studios and Ninja Theory | Gameplay generation research model, February 2025 |
The video platform's rebrand to Skiv carried an embed URL migration deadline of 10 March 2026, so anyone with old muse.ai embeds in production has already had to deal with it. Meta subsequently took over the vacated domain as the web front end for its agent. There is no corporate relationship between the two companies, and there is none between Meta and Microsoft's Muse either.
If you searched "muse spark api" and landed on video hosting documentation, this is why.
Does Meta Offer Muse Spark Credits for Startups?
No. Meta runs no dedicated Muse or Muse Spark startup credit program. Its developer programs page lists only Meta Horizon Start and the Meta Horizon Creator Program, both of which are VR and MR focused and unrelated to the Model API.
This is worth stating plainly because several third-party blogs claim otherwise. Two specific claims circulating are not supported by Meta's own documentation:
"$20 in free Meta Model API credits for new accounts." This appears in third-party write-ups and does not appear on Meta's official pricing page. Treat it as unverified until Meta publishes it.
"The Llama Startup Program covers Muse Spark." The Llama program exists and offers a monthly credit allowance, but no source confirms whether it extends to Muse Spark or the Meta Model API. Meta has not clarified this either way. Plan your budget as though it does not.
The practical answer: Muse Spark is a product you pay list price for. That is fine, and it is the reason to be deliberate about which parts of your stack you fund with credits instead.
Most AI-adjacent spend in a startup's first year is not one vendor. It is a model provider, a compute layer, a vector database, observability, transcription, email, analytics and design tools. Several of those categories have generous startup programs, and AI Perks catalogues $7.7M in credits across 194 companies so you can cover the layers that are fundable and pay cash only where you must.
The sequencing that works: fund everything that has a program, then decide whether the remaining paid line item is worth it. Start at getaiperks.com.
Frequently Asked Questions
What is Muse Spark?
Muse Spark is Meta's foundation model, released 8 April 2026 by Meta Superintelligence Labs under chief AI officer Alexandr Wang. Version 1.1 shipped 9 July 2026 alongside the Meta Model API public preview. It is sold self-serve, is compatible with the OpenAI SDK, and sits underneath Meta's wider Muse product family.
How much does the Meta Model API cost?
Muse Spark 1.1, 1.2 and 1.3 cost $1.25 per million input tokens and $4.25 per million output tokens. Muse Image is $0.01 per image and Muse Voice Transcribe is $3.00 per 1,000 minutes. A Contributor tier charges $0.10 and $0.20 per million tokens if you let Meta train on your data.
Is Muse Spark free for developers?
No. The Meta Model API is paid per token at published rates. The consumer Muse app has a free tier, but that is a personal agent subscription and grants no developer quota. Claims of free API credits for new accounts are not on Meta's official pricing page. Funded alternatives are listed at getaiperks.com.
Is muse.ai still the video hosting platform?
No. The independent muse.ai video hosting and search company, founded by Antonio Roldao, rebranded to Skiv on 3 March 2026 with an embed URL migration deadline of 10 March. Meta later took over the vacated domain as the web front end for its consumer agent. The two companies are unrelated.
Does Meta offer free credits for Muse Spark?
Meta has no dedicated Muse credit program. Its developer programs page lists only Meta Horizon Start and the Meta Horizon Creator Program, both VR and MR focused. Whether the separate Llama Startup Program extends to Muse Spark is unconfirmed. Credits from 194 other companies are tracked at getaiperks.com.
Is Meta's Muse the same as Microsoft's Muse?
No. Microsoft announced Muse in February 2025 as a gameplay generation research model, built with Xbox Game Studios and Ninja Theory and also referred to as WHAM. It has no connection to Meta's 2026 Muse agent or to Muse Spark. The shared name is a coincidence, not a partnership.
Pay list price for the model. Get everything else free.