Meta has released Muse Spark 1.1, a major upgrade to its agentic frontier model, and for the first time opened a paid developer API to its most capable systems. Announced on July 9, 2026, the launch pairs a 1-million-token context window with strong computer-use capabilities and aggressive pricing, putting Meta squarely into the same commercial arena as Anthropic and OpenAI. It is the most significant frontier-model release of the week.
A Second Model, a New Business Model
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 bigger shift may be structural rather than technical: alongside the model, Meta launched the Meta Model API in public preview, marking its entry into the paid-API business that has defined rivals. Unlike the openly downloadable Llama family, Muse Spark 1.1 is proprietary and closed-weight, offered free to consumers through the Meta AI app and meta.ai and sold to developers through the new API.
Meta says the model improves across coding, multimodal reasoning, long-context work and tool-based agentic tasks. US developers can access it immediately, with a waitlist for broader availability, and EU access targeted for later this year subject to regulatory review.
The Million-Token Context Window
The headline capability is a 1-million-token context window that the model can actively manage rather than passively hold. According to Meta, Muse Spark 1.1 remembers earlier actions, retrieves information from much earlier in a task and compacts its working memory in a way that preserves the steps needed later. In practical terms, that is enough room to feed the model an entire codebase, a full set of legal contracts or months of operational logs in a single call.
Active context management is an important distinction. A large window is only useful if a model can find and reuse the right details buried inside it; models that simply accept long inputs often lose the thread on complex, multi-step work. Meta's pitch is that Muse Spark 1.1 treats its context as a workspace to be curated, not just a buffer to be filled.
Built for Computer Use
The model is tuned for computer-use workflows that span multiple applications with information changing on the fly. Meta says it maintains context across extended sessions, adapts to evolving requirements and navigates unfamiliar interfaces with minimal human intervention — deciding when to automate a step and when to interact with a UI directly rather than reasoning through every click.
Muse Spark 1.1 also supports advanced agentic orchestration. It can call native tools, connect to MCP servers and use custom skills, and it can act as a primary agent that builds a plan and delegates subtasks to multiple subagents working in parallel. That parallelism is aimed at completing complex workflows faster than a single sequential agent could.
Benchmarks and Honest Caveats
Meta reports state-of-the-art results on several agentic and domain benchmarks, including MCP Atlas, JobBench, Humanity's Last Exam and FinanceBench, and claims the model competes with more expensive frontier systems such as GPT-5.5 and Claude Opus 4.8 on other tests.
The picture is not uniformly strong, however. Independent analysis flags a clear trade-off:
- Strength: focused gains in tool use, computer use, coding and multimodal understanding.
- Weakness: long-horizon agentic work remains weaker than GPT-5.5 and Opus 4.8, meaning very extended, multi-stage tasks may still favor rivals.
That candor matters for buyers deciding where Muse Spark 1.1 fits. It looks best suited to bounded, tool-heavy automation rather than open-ended, hours-long autonomous projects.
Why It Matters
The most disruptive element may be price. Muse Spark 1.1 costs about $1.25 per million input tokens and $4.25 per million output tokens — roughly a quarter of competing frontier models — with $20 in free credits for new accounts. By pairing frontier-class agentic features with steep discounts, Meta is applying direct pricing pressure to a market where premium reasoning models often cost far more.
The release also signals a strategic pivot. For years Meta's identity in AI was open weights and the Llama ecosystem. Launching a closed, paid, proprietary model API alongside free consumer access shows the company hedging toward the commercial model its competitors have proven, while keeping a consumer on-ramp at meta.ai.
The Bigger Picture
Muse Spark 1.1 arrives during an unusually dense stretch of frontier launches, with new model families from multiple labs appearing within days of one another and a growing emphasis on models built for specific jobs, budgets and speeds. Meta's entry sharpens that trend, offering an agent-focused, long-context model at a price designed to win developers deciding which platform to build on. Whether the discount offsets its weaker long-horizon performance will determine how much of the market it captures.
