Meta Superintelligence Labs released Muse Spark 1.1, the company's first paid AI model, at $1.25/$4.25 per million tokens — the lowest frontier-class pricing available. With a 1M token context window and #1 rankings on MCP Atlas, JobBench, and Finance Agent V2, the closed-source release marks Meta's strategic shift from open weights to paid API access.
Meta Superintelligence Labs released Muse Spark 1.1 on July 9, and for the first time in the company's AI history, it's not free. The model costs $1.25 per million input tokens and $4.25 per million output — the lowest frontier-class pricing on the market. More importantly, it signals that even the strongest open-source champion has found the limits of giving everything away.
Muse Spark 1.1 is a multimodal reasoning model with a 1 million token context window, built explicitly for agentic tasks. It claims the number one spot on MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2 — agent-centric benchmarks that measure practical task completion rather than academic reasoning. On the held-back Harvey legal-agent benchmark, it scores 20% against Anthropic Fable's 11%, a gap that suggests real-world legal reasoning may be advancing faster than public benchmarks indicate.
The model ships with computer use across desktop, browser, and mobile. It can delegate to parallel subagents. It integrates with the Meta Model API — the company's first paid developer API, offered in public preview with $20 in free credits, US-only at launch. Replit, Cline, and Box are named as early integration partners.
The pricing is the story. $1.25/$4.25 per million tokens is less than half of GPT-5.6 Terra and roughly a quarter of Opus 4.8. For a frontier-class agentic model with a 1M context window, that's genuinely disruptive. Meta is pricing to acquire market share, not to cover costs — the classic platform play from a company that can afford to lose money on API revenue for years.
The context window deserves attention. 1 million tokens is enough to hold entire codebases, full legal documents, or multi-year financial records in working memory. Combined with the agentic architecture, it means Muse Spark 1.1 can operate on whole systems rather than fragments — analyze an entire repository, review a complete contract, audit a full financial dataset. That capability, at that price point, changes the economics of AI-assisted professional work.
Mark Zuckerberg returned to X to announce the model — his first substantive appearance on the platform since the July 2 town hall where he admitted AI agent development had stalled for four months. The contrast is notable. Two weeks after acknowledging the agent problem, Meta shipped a model purpose-built for agentic tasks and priced to move.
No open weights. This is the first Meta AI model released without public weights, and it marks a significant shift in the company's strategy. Meta built its AI credibility on open-source releases — Llama, Llama 2, Llama 3, Llama 4. Muse Spark 1.1 is closed. The API is paid. The free tier is a credit coupon, not a model download.
The rationale is partly technical — the model is deeply integrated with Meta's infrastructure and tool-use capabilities, making open-weight release less meaningful — and partly commercial. Meta wants developers building on its API, not running the model on their own hardware. The shift from open weights to paid API access mirrors the broader industry trend: the most capable models are increasingly delivered as services, not artifacts.
What does this mean for the competitive landscape? Four frontier agentic models now compete directly: GPT-5.6 Sol, Claude Opus 4.8, Grok 4.5, and Muse Spark 1.1. Each takes a different approach to pricing, access, and openness. Meta's entry with the lowest price and the longest context window puts pressure on every other player. The agentic AI market is no longer a two-horse race. It's a brawl.
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
Agentic AI Agentic AI refers to AI systems that can autonomously plan, execute multi-step tasks, use tools, and make decisions with minimal human oversight.
AI Agent An autonomous AI system that can perceive its environment, make decisions, and take actions to achieve goals.
Anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei.
Attention A mechanism that lets neural networks focus on the most relevant parts of their input when producing output.