Meta Superintelligence Labs (MSL) officially released Muse Spark 1.1 on July 9; Alexandr Wang, head of Meta AI, said in a CNBC interview that Muse Spark 1.1 is Meta’s “strongest agent task and coding work model” to date.
Muse Spark 1.1 Core Specs: 1 Million Token Context
According to Meta’s official description, the main specs and features of Muse Spark 1.1 are as follows:
Context window: 1 million tokens, enabling easy retrieval of past records, active memory, and compression of key steps
Multi-agent collaboration: can act as a global “master agent” (collecting context, formulating plans, delegating to sub-agents) or focus on a single duty “sub-agent” (intelligently deciding when to escalate and return)
Computer Use: executes dynamic long workflows across multiple applications; can navigate unfamiliar UIs with minimal human intervention. Example: when planning a dinner and ordering experiences a sudden change, it can proactively detect it and update decisions immediately
Advanced coding: diagnoses complex bugs, implements enterprise-level new features, and performs large-scale code migrations; widely used by Meta engineers and researchers
Multimodal vision: visual-to-code generation, and extremely detailed image and video descriptions. Facebook Marketplace example: using only a phone video, it can automatically extract product photos and complete the listing
Pricing strategy: $1.25 in, $4.25 out per million tokens
According to Alexandr Wang’s remarks on CNBC, Muse Spark 1.1 API pricing is: $1.25 per million tokens for input and $4.25 per million tokens for output; each new API account receives a $20 free allowance.
Alexandr Wang said the pricing is “very aggressive and appealing” compared with similar products from labs like Anthropic and OpenAI, aiming to maximize adoption to support large-scale usage growth.
Currently, the API is only available for public preview on the Meta developer portal and has not been opened to third-party platforms like OpenRouter; new users can join a waitlist to gradually gain access.
Meta shifts from open source to proprietary commercialization, with Watermelon still in training
According to reports, Meta previously focused its AI strategy on releasing Llama-series models to the open-source community, but is now shifting more toward selling access to proprietary AI models. Alexandr Wang said Meta remains “committed to open source,” and MSL is developing an open-source variant of Muse Spark, but did not disclose a release timeline.
Alexandr Wang also revealed that Meta is training a stronger model codenamed “Watermelon” (Muse Spark’s internal codename is Avocado), but did not say when it will be released.
Also released this week is Muse Image (an image generation model, formerly codenamed Mango), aiming to serve creators and advertisers.
FAQ
What is the Muse Spark 1.1 API pricing, and how can it be accessed?
According to Alexandr Wang, pricing is $1.25 per million tokens for input and $4.25 per million tokens for output, with each new API account receiving a $20 free allowance. Developers can access it via the public preview on the Meta developer portal, and new users can join a waitlist. General users can use it in the “Thinking” mode of the Meta AI App and on the meta.ai website, but it is currently not open to third-party platforms like OpenRouter.
How does Muse Spark 1.1 differ from the original Muse Spark?
According to reports, the original Muse Spark was launched in April 2026 and provided private API previews only to selected partners. Muse Spark 1.1 has been further opened to the public preview on the developer portal, while enhancing agent collaboration, computer use, and coding capabilities, and supporting a 1 million token context window.
What is Meta’s next more powerful AI model?
According to Alexandr Wang’s disclosure, Meta is training a stronger model codenamed “Watermelon,” but he did not specify a release date. An open-source variant of Muse Spark is also planned, but no specific schedule has been provided either.
