Meta has released Muse Spark 1.1, a vision-language model trained for agentic tasks, and opened the Meta Model API, the company’s first paid access point to its models.
Capabilities of Muse Spark 1.1
Muse Spark 1.1 accepts text, images, and video as input and supports very large contexts (up to 1,048,576 tokens). Text output can be up to 131,072 tokens, with a throughput of up to 119 tokens per second. The model supports tool use, prompt caching, and adjustable reasoning levels: none, minimal, low, medium, high, and xhigh.
Agentic operation and orchestration
Meta disclosed only limited technical details about how Muse Spark 1.1 was built, but says the training emphasized context management, operating computers, and coordinating other agents. The model was trained to work inside multiple agentic coding harnesses—programs that connect a model to files, tools, and tests—and to support planning and subagent orchestration.
Muse Spark 1.1 can lead tasks by splitting work across subagents that run simultaneously, shortening total completion time, and it can also operate as a subordinate agent, returning control when a decision would exceed its permissions. Meta states that the model itself, rather than external software, adjusts its input context over long jobs by retrieving earlier details and condensing exchanges while preserving what later steps will need.
For computer control, the model was trained to choose between automating (writing scripts) and acting directly (clicking or typing on a virtual keyboard), opting for whichever approach is faster or simpler. It can also issue multiple actions at once rather than a single action per turn, improving speed and performance.
Performance on benchmarks
Muse Spark 1.1 stands out on benchmarks that measure agentic tool use, though it generally ranks a tier below the highest-performing models in overall intelligence metrics.
-
Artificial Analysis Intelligence Index: With xhigh reasoning, Muse Spark 1.1 scored 51 points, tying Z.ai’s GLM-5.2 and GPT-5.6 Luna (both at max reasoning) and trailing Claude Sonnet 5 (53 points). On this benchmark Muse Spark 1.1 cost $0.26 per Intelligence Index task, lower than any other model at or above its score except GPT-5.6 Luna ($0.21 per task).
-
Arena.ai Text Arena: In blind human head-to-head comparisons, Muse Spark 1.1 placed sixth with a 1,490 Elo rating, behind Claude Fable 5 (1,505 Elo) and several Claude Opus variants, and just ahead of GPT-5.6 Sol set to xhigh reasoning (1,486 Elo).
-
Artificial Analysis Coding Agent Index: At xhigh reasoning Muse Spark 1.1 achieved 71.3 points, just behind GPT-5.6 Luna (71.4) and ahead of Gemini 3.5 Flash (70.1). It cost roughly $1.40 per coding task, among the lowest of the coding agents compared, though it typically takes more time per task than some competitors.
-
MCP Atlas (Scale AI): Muse Spark 1.1 achieved an 88.1% pass rate on multi-step tool use via Model Context Protocol servers, outperforming Gemini-3.5-Flash at high reasoning (83.6%) and Claude Fable 5 (83.3%).
-
JobBench: On professional white-collar tasks, Muse Spark 1.1 scored 54.7%, placing second behind Claude Fable 5 (57.4%) and ahead of Claude Opus 4.8 (48.4%).
Availability and pricing
Muse Spark 1.1 is available for free through the Meta AI app and at meta.ai. API access is provided via the Meta Model API public preview, currently available only in the United States on a waitlist; new users receive $20 in initial credits.
Meta Model API token pricing:
- Input tokens: $1.25 per million
- Cached tokens: $0.15 per million
- Output tokens: $4.25 per million
- Websearch: $2.50 per 1,000 queries
Meta’s output token price is a small fraction of what some competitors charge (the report cites $25–$50 per million for models such as Claude Opus 4.8, GPT-5.6 Sol, and Claude Fable 5), and per-task comparisons indicate the lower price translates into materially cheaper execution for many tasks.
What Meta did not disclose
Meta did not reveal Muse Spark 1.1’s parameter count, architectural specifics, training data, or full training methodology. The company described training goals and behavioral capabilities but provided few implementation details.
Industry context and significance
Muse Spark 1.1 was released amid a flurry of model announcements: on the same day OpenAI opened its GPT-5.6 family to the public; Grok 4.5 had launched a day earlier emphasizing low cost per task; Google later announced Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Meta also introduced Muse Image and announced Muse Video in the same week. Muse Spark 1.1 is the first model accessible via Meta’s new Model API.
Token prices strongly influence which applications are economical at scale. By offering much lower output-token and per-task costs, Meta has altered that gate. Mark Zuckerberg framed Meta’s pricing as a challenge to competitors’ economics, saying other labs’ pricing "is very extreme and has very high margins." Like Google, Meta can offset development and inference costs with advertising revenue, a structural advantage over labs that rely mainly on API fees and subscriptions. If competitors respond by lowering prices, the cost of running agentic systems could fall industry-wide.
Implications for developers
Meta’s approach continues a trend in which behaviors previously implemented as external scaffolding—instruction following, tool use, delegation, context management—are absorbed into model weights through training. As models internalize more of this functionality, developers will need to build less external orchestration by hand, which could accelerate the spread of agentic applications.
In short, Muse Spark 1.1 offers strong agentic tool-use capabilities at lower operational cost, while Meta withholds many technical details. The broader impact will depend on how competitors react on price and how developers adopt the model’s agentic features.
