Meta’s pivot: from open weights to paid, agentic API
Meta Muse Spark 1.1 is a new multimodal reasoning model exposed through the Meta Muse Spark API that targets developers building agentic AI systems, combining long‑context tool use, coding skills, and multimodal understanding with an aggressively priced, fully hosted inference platform.
Meta is rolling out Muse Spark 1.1 from its Superintelligence Labs together with a public preview of the Meta Model API, after the first Muse Spark stayed locked behind a closed partner program and private endpoints. This is not a small iteration; it is Meta’s opening shot at owning agentic AI development instead of merely seeding Llama weights into the ecosystem. The timing is not accidental. The launch comes three months after the first Muse model and against record capital expenditure on compute and data centers, while investors demand a clear AI monetization story after punishing Meta’s stock over an uncertain plan. In effect, Meta is telling developers and Wall Street the same thing: the era of “free” frontier models is over; the platform era starts with Muse Spark 1.1.
What Muse Spark 1.1 can do that Llama and friends cannot
Muse Spark 1.1 is built for long, tool-heavy work rather than chatty demos. Meta says the model can handle a 1‑million‑token context window, retrieve information from far earlier in a task, and compact that context so later steps retain the important details. That matters for production agents that live inside sprawling enterprise workflows: they must remember dozens of API calls, documents, UI states, and partial plans without collapsing under their own history.
On top of that, Meta claims “state-of-the-art or very close to it” agentic reasoning and tool use, with strong gains in computer use, coding, and multimodal understanding over the first Muse Spark. The model can plan work, call tools, operate across external apps and services, speak to MCP servers and custom skills, and coordinate parallel sub‑agents. Its multimodal side is not decorative either: Muse Spark 1.1 supports visual‑to‑code workflows, image and video captioning, and jobs that combine visual or audio inspection with actions on the user’s behalf. In other words, this is tuned around agentic AI development rather than generic question answering.
Pricing: Meta Muse Spark API undercuts frontier rivals
The bolder move is not the model itself but how Meta is selling it. Muse Spark 1.1 ships with a paid developer tier, the first time Meta has attached a price tag to any of its AI models, turning the Meta Muse Spark API into a direct revenue line. The company is offering new accounts USD 20 (approx. RM92) in free credits and then charging USD 1.25 (approx. RM5.75) per million input tokens and USD 4.25 (approx. RM19.55) per million output tokens.
According to Mark Zuckerberg, “The pricing from some of the other labs is very extreme and has very high margins.” Meta says its Meta Model API comes in at about a quarter of what OpenAI and Anthropic charge for their top‑tier models, a very deliberate AI model pricing comparison that positions Muse Spark 1.1 as an OpenAI Anthropic alternative. This is not altruism; it is platform warfare. By undercutting rivals while keeping distribution on its own servers, Meta is steering developers away from downloading weights and toward a dependency on Meta-hosted inference, with lower operational overhead but far tighter platform lock‑in.
Why agentic AI is Meta’s chosen battleground
Meta is not trying to win the “smartest chatbot” crown; it is chasing the automation stack that will sit inside products, IDEs, and enterprise back offices. The new model’s standout improvement is its agentic capabilities—systems that can autonomously complete multistep tasks on a user’s behalf. Wang’s team weighted coding performance heavily, arguing that a model fluent in code is better at juggling parallel workstreams and chaining tools, the core skills an AI agent needs.
Production agents now spend less time generating text and more time calling APIs, writing code, coordinating workflows, and interacting with external tools. That aligns with Muse Spark 1.1’s design for large‑codebase tasks like bug fixing, feature implementation, migrations, automated screenshots, debugging, and validation loops, all wrapped in planning mode, goal conditioning, sub‑agent delegation, and context compaction. Meta even claims it outperforms Google’s latest models on internal benchmarks that cover agentic workflows, coding, and multimodal reasoning. The message to developers is blunt: if you care about agents that get real work done across apps, not just text, Meta wants your workloads.
What this means for developers and the AI platform race
For developers, the public preview of the Meta Model API means you can now call Muse Spark 1.1 directly, with support for tool calling, function calling, and rich developer prompts, and a new portal and waitlist already open. Meta AI itself is now powered by Muse Spark 1.1, and Meta is one of the few companies able to scale such a model to billions of users quickly. For now, the company is betting that its aggressive price point will keep high‑quality intelligence accessible for free consumer tools while carving out a lucrative chunk of the developer market.
At the same time, an open‑source version of Muse Spark is planned, suggesting Meta wants to straddle both hosted and downloadable models instead of picking a single ideology. The rapid follow‑on launch of Muse Image, designed for creators and advertisers, hints at a broader Muse suite where text, code, and media models share a common API surface. The real question is whether developers accept the trade: cheaper, powerful agentic AI development on Meta’s terms, or a more fragmented stack built from open weights and third‑party providers. With Muse Spark 1.1, Meta has stopped preaching openness and started competing as a full‑blown platform.
