Meta Prices 1M-Token Muse Spark 1.1 at $1.25 Per Million Input

Meta has introduced its first paid developer interface with the release of Muse Spark 1.1 and the Meta Model API. Developed by Meta Superintelligence Labs under Chief AI Officer Alexandr Wang, the model targets enterprise coding tasks and long-horizon agentic workloads. This launch marks a significant shift from Meta’s strictly open-weight strategy, introducing a commercial endpoint optimized for inference-time compute scaling and orchestration.

Architecture and Agent Capabilities

Muse Spark 1.1 operates as a multimodal orchestrator relying on hierarchical subagent delegation. The architecture allows a main agent to maintain state and context while spinning up specialized subagents for discrete tasks. If you implement multi-agent coordination patterns, this native hierarchy reduces the scaffolding required at the application layer.

The model features a documented 1,048,576-token context window. To prevent state degradation during long-running tasks, Meta implemented active context compaction and thought compression. This allows the model to process text, images, video, and documents continuously without losing focus on the primary objective.

Muse Spark 1.1 also features native computer use. The model interprets screen pixels directly and outputs mouse and keyboard actions within a sandboxed environment. A dedicated “Thinking” mode scales compute during inference, enabling the model to deliberate on complex multi-step GUI navigations or codebase refactors before outputting a response.

Benchmark Performance

Meta evaluated the model against frontier offerings, focusing heavily on tool-oriented evaluations. Muse Spark 1.1 leads in professional tool use and agentic coding, though it slightly trails Anthropic in raw computer use execution.

Benchmark	Muse Spark 1.1	GPT-5.5	Claude Opus 4.8
JobBench (Professional Tool Use)	54.7	38.3	48.4
MCP Atlas (Scaled Tool Use)	88.1	Not specified	Not specified
OSWorld-Verified (Computer Use)	80.8	Not specified	83.4
SWE-Bench Pro	61.5	Not specified	Not specified
Terminal-Bench 2.1	80.0	Not specified	Not specified

Meta Model API Pricing

The public preview of the Meta Model API provides OpenAI-compatible endpoints with support for structured outputs and parallel tool calling. Meta designed the pricing specifically to undercut incumbent API providers for high-volume agentic tasks.

Input tokens cost $1.25 per million. Output tokens cost $4.25 per million. By comparison, GPT-5.5 costs $5.00 for input and $30.00 for output, while Claude Opus 4.8 costs $5.00 for input and $25.00 for output. This makes Meta’s endpoint roughly 75 to 85 percent cheaper for equivalent workload processing. If you need to reduce LLM API costs in production, this price point changes the financial viability of running continuous background agents.

Safety Framework and Availability

The Meta Model API is currently in public preview for US developers, and new accounts receive $20 in promotional credits. There is no EU availability at this time.

Prior to release, Meta evaluated the model against its Advanced AI Scaling Framework. Pre-mitigation testing flagged high-risk capabilities in the chemical, biological, and cybersecurity domains. Meta applied multi-layered mitigations to these vectors, reducing the residual risk to moderate or lower levels before clearing the model for deployment.

If you build UIs that rely heavily on parallel tool calling or hierarchical agent delegation, you can begin testing Muse Spark 1.1’s inference-time compute scaling for free via the Meta AI app before migrating production workloads to the metered API.