In This Article
Key Takeaways
- On July 9, 2026, Meta released Muse Spark 1.1 and opened the Meta Model API in public preview — its first paid model API, putting it in the same business as Anthropic and OpenAI.
- Reported pricing is $1.25 per million input tokens and $4.25 per million output tokens, which Meta framed as roughly a quarter of comparable competitor rates.
- The model has a 1M-token context window, tool calling, MCP server support, computer use, and multimodal understanding across images, video, and audio.
- For builders, the interesting question is not the benchmark but the price-per-capability: a cheap, long-context, tool-using model changes what is affordable to build.
Meta just did something it had never done before: it put a price tag on one of its models. On July 9, 2026, Meta released Muse Spark 1.1 and opened the Meta Model API in public preview. For a company that built its AI reputation on releasing open weights for free, charging for API access is a genuine strategic turn — and the number it chose is the most interesting part.
Muse Spark 1.1 is the second model from Meta Superintelligence Labs and an upgrade to the original Muse Spark that debuted in April. The upgrade adds capability, but the release is really about the business model: Meta is now selling inference the way Anthropic and OpenAI do. Let us look at what it costs, what it does, and whether it belongs in your stack.
What Meta shipped
Two things landed together. The first is the model itself, Muse Spark 1.1, described by Meta as a complete agentic foundation — a multimodal reasoning model built for tasks that involve calling tools and taking actions, not just answering questions. The second is the delivery vehicle: the Meta Model API, now in public preview, through which developers can call the model on a pay-per-token basis for the first time.
That combination is the news. Meta has shipped strong models before and given the weights away. Selling hosted API access is new, and it signals that Meta wants a seat at the table where the actual money in this market changes hands — the metered inference business.
The price is the headline
According to multiple reports of the launch, Muse Spark 1.1 is priced at $1.25 per million input tokens and $4.25 per million output tokens. Meta positioned that at roughly 25% of what Anthropic and OpenAI charge for comparable models. That is aggressive, and it is aggressive on purpose.
Undercutting the incumbents on price is the classic move of a well-capitalized latecomer. Meta can afford to sell inference near cost because API revenue is not why Meta exists; for Anthropic and OpenAI, that revenue is the business. When a company that does not need the margin enters a market where everyone else does, prices tend to move. If you buy tokens at scale, this is the kind of launch that gives you leverage in your next renewal conversation, whether or not you ever ship a line of Muse Spark in production.
Muse Spark 1.1 at a glance
| Attribute | Muse Spark 1.1 |
|---|---|
| Released | July 9, 2026 |
| Input price | $1.25 / million tokens (reported) |
| Output price | $4.25 / million tokens (reported) |
| Context window | 1,000,000 tokens |
| Tooling | Tool calling, MCP servers, custom skills |
| Modalities | Text, images, video, audio |
What the model actually does
Price only matters if the capability is real, so here is what Meta says the model brings. It carries a 1 million token context window and is built to actively manage that context across long tasks — the same order of magnitude we covered in our explainer on million-token context windows. On tooling, Meta says the model zero-shot generalizes to new native tools, MCP servers, and custom skills. That MCP support matters: it means Muse Spark can plug into the same tool ecosystem that Claude, Google's agents, and others already speak, which lowers the switching cost of trying it. If you are new to the protocol, our MCP servers list for 2026 is a good starting map.
Beyond tools, Meta describes strong computer-use ability — workflows that unfold across multiple applications, with the model deciding when to automate and when to drive the interface directly — plus multimodal understanding across images, video, and audio, with particular strength in turning visual input into code. In short, it is pitched as an agent model, not a chatbot: something you point at a multi-step job with tools attached, not something you only ask questions.
One honest caveat on the numbers
Meta's own launch post announces the public preview and the model's capabilities but does not publish the per-token price directly. The $1.25/$4.25 figures come from consistent reporting across multiple outlets covering the launch, and Meta's framing of "about a quarter of competitor pricing" is widely quoted. Treat the exact numbers as strong secondhand reporting rather than a figure lifted from Meta's price page, and confirm current rates in the Meta Model API console before you build a budget on them.
Why Meta is charging at all
The strategic logic is worth naming because it explains the price. Meta's advertising business does not need API revenue, so Meta can treat inference as close to a commodity and price it to win developers rather than to earn a margin. Cheap, capable models attract builders; builders create an ecosystem; an ecosystem gives Meta distribution and data that feed back into its core business. Selling tokens is a means, not the end.
This is the same dynamic behind Meta's enormous infrastructure spending, which we covered in our piece on Meta's AI capital expenditure. A company pouring that much into compute wants that compute utilized, and a low-priced public API is one way to fill it. For everyone buying tokens, the practical result is downward pressure on prices across the whole market — a good thing if you are the one paying the bill.
What it means for builders
Reframe the launch as a price-per-capability event rather than a leaderboard event. A model with a million-token context, tool calling, MCP, and computer use, priced at roughly a quarter of the incumbents, changes the arithmetic on things that were previously too expensive to run at scale — large-document analysis, long agent loops that burn a lot of tokens, batch processing over big corpora. Workloads you shelved because the token cost did not pencil out deserve a second look.
The counterweight is maturity. Anthropic and OpenAI have years of tooling, reliability data, and ecosystem around their APIs. A public preview is exactly that — a preview — and preview reliability, rate limits, and support are unknowns you should test rather than assume. The right posture is neither "switch everything" nor "ignore it," but "run a real evaluation on a workload you actually care about and let the numbers decide."
How to evaluate it
Three steps. First, pick one real workload — not a toy prompt — that is either cost-sensitive or context-hungry, since those are where Muse Spark's price and window most plausibly help. Second, run the same workload on your current model and on Muse Spark 1.1 and compare quality, latency, and cost head to head; do not trust a benchmark to stand in for your task. Third, price both options out at your real volume before deciding — our LLM price calculator and API cost estimator let you model input and output tokens at scale so the comparison reflects your actual bill, not a single call.
If it wins on your workload, you have a cheaper option and a genuine second source. If it loses, you have learned exactly where it falls short and gained a data point for your next negotiation with your current provider. Either outcome is useful, which is why a launch like this is worth an afternoon of testing even if you never adopt it.
Price your AI workloads before you commit
Our free calculators let you model token costs across providers and see what a switch would actually save — no signup, no sales call.
Open the LLM Price CalculatorSources: Meta AI, “Introducing Muse Spark 1.1”; AI Weekly on Muse Spark pricing; Tech Times, “Meta’s Muse Spark 1.1 Opens Paid API”. Per-token prices are as reported by third-party outlets; confirm current rates in the Meta Model API console. Analysis by Precision AI Academy.
Common questions
How much does Muse Spark 1.1 cost? Reported pricing is $1.25 per million input tokens and $4.25 per million output tokens, positioned at roughly a quarter of comparable Anthropic and OpenAI rates. Confirm current numbers in Meta's console.
What is the context window? One million tokens, with the model built to actively manage that context across long agentic tasks.
Does it support MCP and tool calling? Yes. Meta says the model zero-shot generalizes to new native tools, MCP servers, and custom skills, and handles computer-use workflows across multiple applications.
Is this open weights? No — this launch is about the paid Meta Model API in public preview. It marks Meta charging for hosted access, a change from its earlier open-weight releases.