Meta opens Muse Spark to developers with a low-priced coding model and 1M context window

Why it matters

Meta is using price to force Muse Spark into the model-buying conversation, but API adoption will depend on the operational details Zuckerberg did not disclose.

Mark Zuckerberg (@finkd) said Meta is releasing Muse Spark 1.1, a new version of its agentic and coding model, and making it available to developers through the new Meta Model API for the first time.

The release gives Meta a cleaner answer to a problem that has followed its AI push all year: Muse Spark has been built directly into Meta AI and the company's consumer surfaces, but outside developers have had limited access to the underlying model. Zuckerberg framed the July 9th launch around price and agentic performance, writing in a three-post thread that Muse Spark 1.1 is available in the Meta Model API and in Meta AI, with a focus on "strong agentic and multimodal models at very low cost."

The price is the part Meta wants developers to notice first. Meta told Axios it is charging $1.25 per million input tokens and $4.25 per million output tokens for Muse Spark 1.1, a rate Axios described as cheaper than the new Grok 4.5 and below Anthropic's Opus model. That is Meta's most direct attempt yet to compete for API traffic on cost, rather than relying only on the distribution advantage of Facebook, Instagram, WhatsApp, Messenger, Threads and Meta AI.

Zuckerberg said Muse Spark 1.1 is strongest at "agentic performance, tool use, and computer use." The concrete technical claims are aimed at the workloads that have become the buying committee for frontier models: long-running coding tasks, tool-calling agents, browser automation and workflows that break a job into multiple parallel steps. According to Zuckerberg, the model supports a 1M-token context window, can delegate execution to sub-agents running in parallel and is trained to use computer interfaces on desktop, mobile and browser.

Meta did not disclose model size, parameter count, rate limits or third-party benchmark scores in Zuckerberg's thread. That leaves developers with a familiar tradeoff: Meta is offering a low published token price and access to a model already being deployed across a massive consumer network, while keeping the technical disclosure thinner than the rate card. For startups choosing a model provider, price per token is only one part of the real cost. Latency, reliability, context-window behavior, tool-call accuracy, data policies and rate limits will decide whether Muse Spark 1.1 becomes a production default or an experiment teams keep behind an abstraction layer.

The release also marks a shift from Meta's April 8th Muse Spark launch. In that announcement, Meta said Muse Spark was the first large language model from Meta Superintelligence Labs and that it was powering the Meta AI app and meta.ai, with private API preview planned for select partners. Meta also said then that it hoped to open-source future versions of the model, a notable line because Muse Spark itself arrived as a proprietary system rather than a continuation of Meta's open-weight Llama posture.

Zuckerberg's founder imprint is all over the product strategy. Meta's AI work has moved from the Llama-era developer narrative toward a more controlled assistant strategy tied to the company's own surfaces and user context. The April Muse Spark announcement said Meta Superintelligence Labs had rebuilt the company's AI stack over nine months and positioned Muse Spark as the first step toward Zuckerberg's "personal superintelligence" thesis: an assistant grounded in the relationships, recommendations and public content already flowing through Meta's apps.

That strategy depends on more than chat. On May 12th, Meta expanded Muse Spark across voice conversations, shopping features and glasses. The company said Muse Spark was starting to roll out to Ray-Ban Meta and Oakley Meta glasses in the U.S. and Canada, with Meta Ray-Ban Display planned for the summer, and across WhatsApp, Instagram, Facebook, Messenger and Threads. In practice, that makes Muse Spark both an API model and the reasoning layer for consumer product features where Meta has data and distribution that model-only labs do not.

The July 9th release follows another Meta Superintelligence Labs launch earlier this week. On July 7th, Meta introduced Muse Image and previewed Muse Video, saying Muse Image integrates with Muse Spark and can use search and coding tools for image generation. That matters because Meta is building the Muse line around tool use across modalities, rather than treating text, image and video models as separate products. Muse Spark 1.1 is the developer-facing move in that plan: give builders a low-cost model for coding and agents, while Meta uses the same family of systems inside its own apps.

Alexandr Wang is the other executive central to the rollout. Axios reported that Wang, brought in to reboot Meta's AI effort, said improving coding and agentic tasks was a priority for this release. Axios also reported that Meta is using Muse Spark 1.1 to power the thinking mode in the Meta AI app and website, and that a larger model, code-named Watermelon, remains in training for release later this year.

The safety backdrop is already documented for the original Muse Spark line. In a June 2026 safety and preparedness report posted to arXiv, Meta researchers wrote that Muse Spark was evaluated under Meta's Advanced AI Scaling Framework across chemical and biological, cybersecurity and loss-of-control risks. The report said pre-mitigation chemical and biological capabilities likely reached Meta's "high risk" category before safeguards, and that Meta implemented mitigations before release. That is relevant because Muse Spark 1.1 is explicitly pitched for agents and computer use, the same direction that can make models more useful to developers and harder to contain when they are connected to tools.

For Meta, the Model API turns Muse Spark from a consumer assistant upgrade into a platform bet. Zuckerberg's thread makes the commercial target clear: win developer workloads where OpenAI, Anthropic, Google and xAI already sell high-capability models, then use Meta's cost structure and distribution to pressure the market. The unanswered question is whether developers will trust a Meta-controlled proprietary model for core application logic, especially when the company has not yet supplied the full technical and operational detail that serious buyers use before moving traffic.