Muse Spark artificial intelligence model now carries a price tag, and developers must pay for access. Meta released the upgraded version on Thursday, opening a public preview to United States developers. The company charges $1.25 per million input tokens and $4.25 per million output tokens. Every new account receives twenty dollars in free credits before pay-as-you-go billing starts. You can read this move as Meta finally selling access instead of giving models away.
Alexandr Wang leads the effort, and he calls the pricing aggressive next to rival lab offerings. His team built the model to handle coding work and long chains of agent tasks. Meta Superintelligence Labs trained it on real-world software problems across large enterprise code bases. Wang told CNBC the update marks the best coding and agent performance Meta has shipped. Rivals now face a cheaper option built by a company with enormous computing capacity.
The Meta Model API sits at the center of this shift toward paid developer access. Developers sign up through a portal, test prompts, compare outputs, and prototype their own integrations. Meta limits access to its own properties for now, skipping third-party marketplaces like OpenRouter. Some early partners already hold API keys, and new users enter a waitlist for entry. Replit, Cline, and Box rank among the first companies building on the new system.
Muse Spark artificial intelligence model sets a new price floor
Muse Spark 1.1 pricing lands below Anthropic’s Claude Sonnet 4.6 on both input and output. The rate still runs above cheaper tiers such as GPT 5 mini and Claude Haiku 4.5. Zuckerberg framed the cost as one of the lowest available to developers right now. In my assessment, price alone will not decide which lab wins the coding market. Quality, reliability, and developer trust matter as much as the number on the invoice.
Meta claims strong benchmark results, including wins over Google’s Gemini 3.1 Pro in some areas. The AI coding model handles bug diagnosis, feature builds, and large-scale code migrations. It supports a context window of one million tokens for long-running technical sessions. Engineers can run it as a lead agent or as a subagent inside larger systems. Mark Zuckerberg said, “Muse Spark 1.1 is strongest at agentic performance, tool use, and computer use.”
Wall Street keeps pressing Mark Zuckerberg for returns on enormous artificial intelligence spending commitments. The company spends like its hyperscaler peers, yet it owns no cloud infrastructure business. Meta plans to launch one, and paid model access opens a second revenue line. Earlier Llama releases went to the open source community without any charge to users. Wang says an open source variant remains in development, though he gave no release date.
What Muse Spark artificial intelligence model means for your stack
Muse Spark artificial intelligence model gives you another vendor inside a crowded developer market. Meta trained the release to work with popular agent harnesses developers already run daily. Wang points to health research as one use case, from web searches to academic papers. Your team should test output quality against cost before moving any production workload over. Meta faces a hard climb, and the Muse Spark artificial intelligence model carries heavy expectations. Developers now decide whether the Muse Spark artificial intelligence model earns a permanent slot.
