Meta has released Muse Spark 1.1, the latest iteration of its multimodal AI model developed under the newly formed Meta Superintelligence Labs. The update follows the initial Muse Spark announcement earlier this year and aims to close the gap with leading systems from OpenAI and Anthropic.
Muse Spark 1.1 brings notable enhancements in agentic capabilities, computer use, coding, and multimodal processing. According to Meta, the model can interface with native tools, MCP servers, and custom skills, while also functioning as a central agent that devises plans and distributes subtasks to parallel agents. It features a one-million-token context window, a substantial increase over prior Meta models, with claims that it can actively manage extended context during prolonged interactions. The company positions it as particularly capable for tasks involving direct computer interaction.
For developers, this marks Meta’s first offering of the model through a public API. New accounts receive $20 in initial credits, after which pricing stands at $1.25 per million input tokens and $4.25 per million output tokens—positioned as more accessible than comparable frontier models. Benchmark results shared by Meta show strong performance on several evaluations, including leading scores on MCP Atlas, JobBench, Humanity’s Last Exam, and FinanceBench, while remaining competitive on others against higher-priced alternatives like GPT-5.5 and Opus-4.8.
Safety evaluations were conducted under Meta’s Advanced AI Scaling Framework, with the company stating that the model stays within acceptable risk thresholds for areas such as chemical and biological threats, cybersecurity, and potential loss of control. For general users, Muse Spark 1.1 is accessible in “Thinking” mode via the Meta AI app and meta.ai platform.
The release reflects Meta’s intensified push into advanced AI following a period of perceived lag behind competitors. By establishing Meta Superintelligence Labs, the company has sought to accelerate development of frontier-level systems. Earlier this year, Muse Spark delivered results that placed it in the conversation with top models, though real-world consistency often tells a more nuanced story than benchmarks alone. Agentic features, in particular, continue to face challenges in reliability across unpredictable workflows, where planning and delegation can falter without careful human guidance.
Pricing and accessibility represent a pragmatic angle in an industry where compute costs have drawn increasing scrutiny. Lower entry points could broaden experimentation, yet questions remain about long-term sustainability, data privacy in tool integrations, and the actual productivity gains once initial novelty fades. Multimodal models have shown promise in handling diverse inputs, but sustained performance in complex, multi-step scenarios still varies widely depending on implementation.
Overall, Muse Spark 1.1 adds another capable contender to the rapidly evolving landscape, underscoring how competition is driving incremental improvements in context handling, tool use, and cost efficiency. Its practical impact will ultimately depend on how well it performs beyond controlled benchmarks in daily professional and creative applications.
