Meta launches Muse Spark 1.1 and its first paid Model API

Meta has released Muse Spark 1.1, a multimodal reasoning model engineered specifically for agentic tasks. Alongside the model, the company has launched a public preview of the new Meta Model API. This marks Meta's first foray into offering a paid developer API for its own frontier models, a significant pivot from its previous open-weights strategy.

The launch adds another layer of complexity to an already crowded week for frontier model announcements. Muse Spark 1.1 is positioned to handle large-scale agentic workloads, complex code migrations and multi-step tool orchestration. The timing of the release, coming just days after GPT-5.6 and Grok 4.5, demonstrates the intensity of competition in the frontier model space.

The Meta Model API is currently available in public preview for developers based in the United States. Access for the European Union has not yet been announced. The pricing structure is highly competitive.

The API charges $1.25 per million input tokens and $4.25 per million output tokens. New accounts receive $20 in free credits to begin testing. This pricing sits between the budget-friendly Luna tier from OpenAI and the mid-range models from other providers.

The API is designed for low-friction adoption, supporting both the OpenAI SDK and the Anthropic Messages format. This dual compatibility is strategically important. Developers can point existing applications at the Meta API with minimal code changes. They can test the model against their current solution without re-architecting their systems.

While Muse Spark 1.1 may not top every raw coding benchmark, it demonstrates exceptional strength in tool use. On the JobBench evaluation, which measures professional tool use, the model scored 54.7. This places it ahead of Opus 4.8 at 48.4 and GPT-5.5 at 38.3.

On the MCP Atlas evaluation, which measures scaled tool use across multiple tools and systems, Muse Spark 1.1 scored 88.1. Opus 4.8 achieved 82.2, and GPT-5.5 achieved 75.3. This is a significant advantage for agentic workflows that require orchestration across multiple systems.

The model is trained to orchestrate multi-agent systems to optimise end-to-end latency. It can gather context, formulate a plan and delegate execution across parallel subagents. When acting as a subagent, it understands available tools and knows when to escalate issues back to the main agent.

Muse Spark 1.1 advances the ability of AI to interact directly with computer interfaces. It understands when to write scripts for faster automation and when to click through an interface directly. The model can maintain context across extended sessions and adapt to changing requirements on the fly.

For project delivery professionals, this capability is significant. Many project management tasks involve navigating multiple systems, extracting data and consolidating it into reports. A model that can automate these interactions could substantially reduce manual effort.

For AEC professionals, the true value of AI often lies in orchestration rather than simple text generation. Project delivery requires querying databases, updating schedules and cross-referencing documentation. We note that a model excelling in tool use is perfectly suited for these complex, multi-system workflows.

The aggressive pricing and drop-in compatibility with existing SDKs lower the barrier for firms looking to test alternative models in their existing agentic setups. A firm currently using OpenAI or Anthropic models can test Muse Spark 1.1 with minimal effort.

The 1 million token context window with active compaction is particularly relevant for long-running project tasks. A model can maintain context across multiple days of work, understanding decisions made earlier in the workflow and how they impact current tasks.

Active compaction means the model itself manages which information to retain and which to discard. This is more sophisticated than simple truncation. The model understands which details are critical for future decisions and preserves them.

Meta reports that Muse Spark 1.1 demonstrates strong resistance to jailbreaks and prompt injection attacks. The model shows lower hallucination rates and reduced sycophancy compared to the original Muse Spark. These improvements are important for enterprise deployment where reliability and predictability are critical.

• Meta's entry into the paid API market introduces aggressive pricing at $1.25/$4.25 per million tokens, positioning the model competitively against established providers.

• The model's superior performance in professional tool use makes it highly effective for orchestrating complex workflows across multiple applications and systems.

• The 1 million token context window with active compaction ensures stability during long-running project tasks that span multiple days.

• Dual compatibility with OpenAI and Anthropic SDKs allows developers to easily test the model within existing infrastructure without re-architecting systems.

• Exceptional performance on MCP Atlas and JobBench evaluations demonstrates strength in multi-tool orchestration, the core requirement for agentic project management systems.

Subscribe to the Project Flux newsletter for a comprehensive breakdown of how API pricing strategy impacts the broader AI ecosystem.

All content reflects our personal views and is not intended as professional advice or to represent any organisation.