Meta Releases Muse Spark 1.1 and Meta Model API

Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model designed for agentic tasks, alongside a public preview of the Meta Model API on an unspecified date. This marks a shift for Meta, as their models will now be accessible via a closed, hosted, and metered API rather than primarily as open weights. Muse Spark 1.1 offers improvements in tool use, computer use, coding, and multimodal understanding compared to its predecessor. The model boasts a context window of 1,000,000 tokens, with API documentation listing 1,048,576 tokens. Its core capabilities include processing text, images, video, and documents as input, with text as output. The API also supports structured output, parallel tool calling, a Files API, and prompt caching. Integrating a web_search tool into Responses API calls enables the retrieval of cited answers.

Access to Muse Spark 1.1 is bifurcated. Consumers can utilize it for free in 'Thinking' mode within the Meta AI app and on meta.ai. Developers will be charged $1.25 per million input tokens and $4.25 per million output tokens. New accounts receive $20 in complimentary credits. The initial launch specifies that the public preview of the Meta Model API is available in the US only, with no current access for the European Union. This API provides a new avenue for developers to integrate Meta's advanced AI capabilities into their applications.

Performance benchmarks published by Meta position Muse Spark 1.1 against other leading models, including GPT-5.5 and Gemini 3.1 Pro. In scaled tool use, Muse Spark 1.1 achieved 88.1%, surpassing GPT-5.5's 82.2% and Gemini 3.1 Pro's 75.3%. For JobBench Professional tool use, it scored 54.7%, ahead of GPT-5.5's 48.4%. On Humanity's Last Exam, a reasoning benchmark, Muse Spark 1.1 reached 62.1%, compared to GPT-5.5's 57.9%. However, in OSWorld-Verified Computer use, GPT-5.5 slightly outperformed Muse Spark 1.1 with 83.4% versus 80.8%. For SWE-Bench Pro real-repo coding, Muse Spark 1.1 scored 61.5%, while GPT-5.5 achieved 69.2%. In DeepSWE 1.1 long-horizon coding, Muse Spark 1.1 scored 53.3%, with GPT-5.5 at 59.0%. For visual reasoning tasks using BabyVision, Muse Spark 1.1 achieved 76.3%, lower than GPT-5.5's 81.2% and Gemini 3.1 Pro's 83.6%.