Meta has launched Muse Spark 1.1, an updated AI model designed for coding and agentic work, positioning itself as a direct competitor to OpenAI and Anthropic in a rapidly consolidating market. The model was introduced on Thursday by Alexandr Wang, Meta's AI chief who leads Meta Superintelligence Labs, following the original Muse Spark release in April that was available only to select partners.

What Can Muse Spark 1.1 Actually Do?

Muse Spark 1.1 is built to handle the kinds of coding tasks that developers spend hours on manually. The model can diagnose and fix bugs, implement new features in large codebases, and run code migrations. Beyond text-based coding, it also supports multimodal tasks, meaning it can process images, video, and audio, and take actions on behalf of users, such as creating a Facebook Marketplace listing from a smartphone video.

What sets this model apart is its design philosophy. Wang said Meta trained the model to work with popular agentic coding tools and harnesses that developers already use. This includes support for planning mode, subagent delegation, and context compaction, features that make it easier to integrate into existing workflows.

The model comes with a million-token context window, which translates to roughly one million words of information it can process at once. This large capacity is particularly valuable for developers working with massive codebases where understanding the full scope of a project is critical.

Why Is Replit's CEO Calling This a Game-Changer?

Early partners have already weighed in on the model's potential. Replit CEO Amjad Masad offered a particularly strong endorsement, calling it "a complete agentic foundation," citing its million-token context window and multimodal support. Replit, a cloud-based IDE (integrated development environment) that lets developers code directly in a browser, has positioned itself as a platform where AI agents can thrive. Masad's endorsement suggests that Meta's model aligns well with how modern development platforms are evolving.

Amjad Masad

"A complete agentic foundation," said Amjad Masad, CEO of Replit, praising the model's million-token context window and multimodal support.

Amjad Masad, CEO at Replit

How to Access Muse Spark 1.1 and What It Costs

• Availability: Meta is making the API available through a public preview developer portal, with new users able to join a waitlist for access.
• Pricing Structure: The model costs $1.25 per million input tokens and $4.25 per million output tokens, which translates to roughly $1.25 per million words processed for input and $4.25 per million words for output.
• Free Credits: Every new account gets $20 in free credits to start, allowing developers to experiment without immediate cost.
• Competitive Positioning: Wang described the pricing as competitive against Anthropic and OpenAI, though Meta is currently limiting API access to its own properties and is not yet listing on third-party platforms like OpenRouter.

Meta is currently limiting API access to its own properties, meaning developers cannot yet access Muse Spark 1.1 through third-party platforms like OpenRouter. This is a strategic choice that keeps the model within Meta's ecosystem for now, though the company has indicated it plans to expand availability.

What Does This Mean for Meta's AI Strategy?

This launch marks a significant shift in Meta's approach to artificial intelligence. The company previously focused on releasing open-source models under the Llama name, prioritizing accessibility and community contribution. Now, Meta is selling access to proprietary models, a move that reflects pressure from Wall Street to demonstrate returns on the company's growing AI investments.

CEO Mark Zuckerberg faces particular pressure to show financial returns on Meta's AI spending. Unlike rivals such as OpenAI and Anthropic, Meta does not have a cloud business, though the company has said it plans to build one. Muse Spark 1.1 appears to be a cornerstone of that strategy.

Wang noted that a future open-source version of Muse Spark is in development, but he gave no release date. This suggests Meta is hedging its bets, maintaining its open-source commitment while simultaneously building a proprietary business around advanced models.

What Else Is Meta Building?

Muse Spark 1.1 is not Meta's only recent AI release. Earlier this week, Meta released Muse Image, a model for generating images aimed at creators and advertisers. The company also confirmed it is training a more powerful model, code-named Watermelon, though no release date was provided.

Wang added a personal note about his own use of the model, saying he has been using it to help with health research, including reading academic papers and accessing personal health data. This suggests the model's capabilities extend beyond traditional coding tasks into knowledge work and research applications.

The launch of Muse Spark 1.1 signals that the AI coding market is becoming increasingly crowded, with major tech companies competing not just on model quality but on integration with existing developer tools and platforms. Replit's early adoption suggests that developers building on cloud-based platforms are ready to incorporate these new models into their workflows.