Muse Spark 1.1 in one sentence: a long-context agent engine that outsiders can finally use
Meta’s Muse Spark 1.1 API is a multimodal reasoning model with a 1‑million‑token context window and public developer access, designed to run long, tool-heavy AI agent workflows, coordinate subagents, and handle visual and audio inputs as part of complex application logic. Meta is rolling out Muse Spark 1.1 from Meta Superintelligence Labs alongside a public preview of the Meta Model API, opening the model beyond Meta AI’s own interface for the first time. That matters more than a routine version bump: it signals that Meta wants Muse Spark to be a platform for AI agents, not just a consumer chatbot. Developers now get the same core engine that powers Meta AI, including tool calling, function calling, and custom developer prompts. In other words, Meta has moved from talking about “agentic workflows” to giving you a concrete, programmable surface to ship them.
Benchmarks, costs, and what Meta’s 51 score really buys you
On paper, the Muse Spark 1.1 API sits squarely in the top tier of mid‑range models. It scores 51 on the Artificial Analysis Intelligence Index, edging past Google’s Gemini 3.5 Flash at 50 and Gemini 3.1 Pro Preview at 46. “Muse Spark 1.1 used 94 million output tokens to run the full Intelligence Index, fewer than GLM‑5.2 (max) at 141 million and GPT‑5.6 Luna (max) at 125 million.” It is also the new state of the art on MedScribe and TaxEval, taking the top spot from Fable 5 while being 10x cheaper and twice as fast on those benchmarks. For agent builders, this isn’t academic. Meta is pricing Muse Spark 1.1 at USD 1.25 (approx. RM5.80) per million input tokens and USD 4.25 (approx. RM19.70) per million output tokens, with cache hits discounted to USD 0.15 (approx. RM0.70) per million. That combination of high benchmark performance and aggressive pricing changes the default calculus of which model you pick for production workloads.
Why Meta’s agentic design matters more than raw IQ points
Muse Spark 1.1 is explicitly tuned as an AI agent engine, not just a generic chat model. Meta says the release brings major gains in agentic workflows, computer use, coding, and multimodal understanding compared with Muse Spark 1.0. The model is built for long, tool‑heavy tasks: it can plan work, call tools, operate across external apps and services, use MCP servers and custom skills, and coordinate parallel subagents. Its 1‑million‑token context window lets it retrieve information from much earlier in a task and compact context so later steps still contain the important details. In practical terms, that means you can run a multi‑hour automation — say, a research agent plus a coding agent plus a QA agent — without constantly micromanaging context or rebuilding state. On Humanity’s Last Exam, Muse Spark 1.1 hits 45%, within a point of Claude Opus 4.8 at 46% and ahead of GPT‑5.5 at 44%, but the more important story is that it keeps this level of reasoning over very long workflows.
Multimodal AI workflows: from visual-to-code to multi-app computer use
Where Muse Spark 1.1 becomes especially interesting is multimodal AI workflows. Meta says the model excels at multi‑app computer‑use workflows by maintaining context across extended sessions and choosing intelligently between scripting, direct UI interaction, and batched actions. That makes it suitable for agents that handle full desktop or browser automation: writing scripts when APIs exist, clicking through when they do not, and batching repetitive steps. The model also adds stronger multimodal workflows, including visual‑to‑code generation, image and video captioning, and tasks that require it to inspect visual or audio inputs while acting on a user’s behalf. In effect, the Muse Spark 1.1 API lets you build agents that can read a product screenshot, generate test code, drive a web interface, and summarize video recordings without constant human intervention. When you pair that with its improved scientific reasoning and coding gains noted by Artificial Analysis, you get a multimodal execution engine that’s much closer to an actual digital worker than a text‑only assistant.
From Llama 4’s missteps to a more transparent, democratized API
Meta’s decision to open the Muse Spark 1.1 API is as much about trust as it is about capability. The company’s last major launch before this run, Llama 4, was a rough chapter after it emerged that benchmark numbers had been assembled using different model versions than what was actually shipped. That episode led to a reshuffling of Meta’s AI organization, a large stake in Scale AI, and Alexandr Wang taking over at Meta Superintelligence Labs. Muse Spark 1.1 is presented as evidence that the rebuild is working, with results independently verified by a third party rather than self‑reported. Crucially, the new API marks a shift from the earlier Muse Spark rollout, which was limited to Meta AI and a private preview. Meta’s evaluation report says Muse Spark 1.1 extends access to external developers via an API that supports tool calling, function calling, and developer prompts. The model is available now through Meta’s own API, with no third‑party hosting confirmed at launch. If you care about democratizing advanced AI, this is a clear signal: Meta is turning frontier‑adjacent capabilities into something agents and multimodal applications can use today, not a year from now.
