The release arrives at a moment when the broader AI industry is racing to build tools that can act on a user's behalf rather than simply respond to prompts. Agentic capability, long context handling, and cost efficiency have become the three metrics developers weigh most heavily when choosing a model provider, and Meta's latest release touches on all three at once.

1. What Is Meta Muse Spark 1.1?

Meta Muse Spark 1.1 is a multimodal reasoning model built by Meta Superintelligence Labs, the research group led by Alexandr Wang. It is designed specifically for agentic tasks such as planning, delegating, and acting across tools and interfaces, rather than simply generating text responses. According to Meta's official announcement, Muse Spark 1.1 represents a significant upgrade from the original Muse Spark model released in April 2026, with major gains in tool use, computer use, coding, and multimodal understanding.

The model is available in two ways. Developers can access it through the newly launched Meta Model API, now in public preview for users in the United States. Everyday users can also try it for free inside the Meta AI app and on meta.ai, where it runs in a "Thinking" mode built for more deliberate, multi step reasoning.

This dual release path matters. It means the model is not locked behind a developer waitlist alone. Consumers get a taste of its reasoning ability through the free app, while businesses and independent developers can build directly against the API once they are ready to integrate it into their own products.

2. Why Meta Muse Spark 1.1 Marks a Structural Shift for Meta

For years, Meta's public AI strategy centered on open weights through the Llama family of models. Anyone could download, fine tune, and deploy Llama models without paying Meta directly for API access. Meta Muse Spark 1.1 changes that pattern. It is closed, hosted, and metered per token, placing it in direct competition with paid offerings from OpenAI and Anthropic.

This is not simply a new model release. It signals that Meta now sees a viable revenue path in charging developers for frontier level capability, while still maintaining a presence in open source through Llama and a promised future open variant of Muse Spark. The launch was significant enough that Mark Zuckerberg posted on X for the first time in three years, describing the model as strong at agentic and coding work and priced very low, according to reporting from TechCrunch.

With hundreds of billions of dollars already committed to AI infrastructure, Meta has faced growing pressure to show a direct return on that spending. A metered API gives the company a new revenue stream while keeping usage tied to its own ecosystem, an approach that differs sharply from the fully open distribution model Meta relied on for years with Llama.

3. Key Capabilities of Meta Muse Spark 1.1

3.1 Agentic Orchestration and Long Context

One of the defining features of the model is its ability to manage complex, multi step projects through multi agent orchestration. As the main agent, Meta Muse Spark 1.1 gathers context, builds a plan, and delegates parts of the task to parallel subagents. As a subagent, it understands its assigned role, knows which tools are available to it, and escalates back to the main agent when needed.

Supporting this is a context window of 1 million tokens, which the model actively manages rather than simply filling up. It retrieves information from much earlier in a session and compacts what it no longer needs, while preserving the steps that matter for later work. This matters because long agentic workflows often break down when a model loses track of earlier decisions, and the model is specifically trained to reduce that failure mode.

3.2 Computer Use

Meta Muse Spark 1.1 is trained to operate real interfaces rather than simulate a text based version of them. It navigates unfamiliar applications, adapts when information changes mid task, and decides for itself whether to automate a step with a script or interact with the interface directly through clicks. In one example Meta shared publicly, the model handled a dinner party ordering task and adjusted its actions on its own when new context appeared partway through, without needing the user to step in.

3.3 Coding Performance

Coding is where the model shows its most significant improvement over its predecessor. Meta reports that Meta Muse Spark 1.1 can diagnose and fix complex bugs, implement new features inside enterprise scale codebases, and carry out large code migrations. It also adapts to popular agentic coding harnesses, supporting planning mode, goal conditioning, subagent delegation, and context compaction, features many development teams already rely on in their existing workflows.

3.4 Multimodal Understanding

Beyond text, Meta Muse Spark 1.1 accepts image, video, audio, and PDF input. Its multimodal strength shows up most clearly in tasks where perception and action need to happen together, such as reviewing a video, extracting the relevant details, and then using those details to complete a task inside a browser or application on the user's behalf.

4. How Does Meta Muse Spark 1.1 Perform on Benchmarks?

Benchmark results for Meta Muse Spark 1.1 show a model that performs unevenly depending on the task type. On Vibe Code Bench v1.1, it scored 72.2, a sharp jump from under 20 on the original Muse Spark released in April. On the Artificial Analysis Coding Agent Index, run through the Opencode harness, it scored 69, close to GPT-5.5 and ahead of Anthropic's Claude Opus 4.8.

On other benchmarks, including SWE-Bench Pro and Terminal-Bench, the model trails both Opus 4.8 and GPT-5.5. Independent evaluators have also noted that real world performance can land below vendor reported figures, a pattern seen across the industry since every lab tends to select benchmark sets that highlight its own strengths. Taken together, the results suggest Meta Muse Spark 1.1 is less a pure coding leader and more an orchestration focused model built for managing multi step agentic work.

This uneven benchmark picture is worth keeping in mind for anyone comparing model providers purely on published scores. A model that leads on one benchmark and trails on another often reflects differences in training priorities rather than a clear overall winner, and developers evaluating this model for a specific use case are generally better served running their own tests against representative tasks before committing to a provider.

5. Safety Measures Behind Meta Muse Spark 1.1

Before release, Meta evaluated the model under its Advanced AI Scaling Framework, which defines evaluation methods, threat models, and deployment thresholds for its most advanced systems. Across the three frontier risk categories the framework tracks (chemical and biological risk, cybersecurity, and loss of control) Meta reports that the model operates within safe margins.

Meta also states the model shows stronger resistance to direct jailbreaks and indirect attacks originating from untrusted data or prompt injection attempts, along with lower hallucination rates and reduced sycophancy compared with the original release. Full technical detail is documented in Meta's published evaluation report, linked from the official Muse Spark 1.1 announcement.

6. Meta Muse Spark 1.1 Pricing: How Much Does It Cost?

Pricing is arguably the clearest statement Meta is making with this release. Meta Muse Spark 1.1 is priced at $1.25 per million input tokens and $4.25 per million output tokens through the Meta Model API. New developer accounts receive $20 in free credits to get started.

For context, that pricing sits slightly above OpenAI's GPT-5 mini and Anthropic's Claude Haiku 4.5, roughly in line with GPT-5.6 Luna, and well below Anthropic's Claude Sonnet 4.6, according to reporting from Reuters. Meta has framed the pricing as aggressive by design, aiming to make it an attractive option for developers running high volume agentic workloads.

7. Availability and Early Adoption

Access to Meta Muse Spark 1.1 is currently limited to developers in the United States through the Meta Model API's public preview. There is no confirmed timeline yet for availability in the European Union, and Meta has noted that current preview pricing could change once the API exits public preview.

Even at this early stage, the model has attracted genuine production interest. Companies including Replit, Cline, and Box are testing it in production workloads, citing its tool use, coding ability, and pricing as reasons for adoption, according to CNBC's coverage of the launch.

8. What Comes After Meta Muse Spark 1.1?

Meta has already confirmed it is training a more capable model, internally code named Watermelon, which is reportedly closing the performance gap with GPT-5.5 on key benchmarks. Meta has also said it plans to release an open source variant of Muse Spark, though no timeline has been announced. Internally, the model is expected to gradually replace the Llama models currently powering AI assistants across WhatsApp, Instagram, and Facebook.

9. Conclusion

Meta Muse Spark 1.1 represents a meaningful shift in how Meta approaches AI, moving from a fully open weights model toward a hybrid strategy that includes a competitively priced, closed API. Its strength lies less in raw coding leadership and more in its ability to plan, delegate, and act across long, multi step agentic workflows.

Whether this pricing advantage converts into lasting developer adoption will depend on how it performs outside Meta's own benchmark suite, in real production environments where reliability matters as much as cost. For now, it adds real pressure on OpenAI and Anthropic to defend their pricing as well as their capability, giving developers one more serious option to evaluate as the agentic AI market continues to move quickly through 2026.