Meta Superintelligence Labs has released Muse Spark 1.1, a multimodal reasoning model purpose-built for agentic tasks, alongside a public preview of the Meta Model API. Unlike Meta’s previous open-weights releases, Muse Spark 1.1 is a closed, hosted model, priced per token.
The model accepts text, images, video, and document inputs and generates text outputs, with a context window of 1,000,000 tokens (1,048,576 per the API documentation). It supports adjustable reasoning effort per request and includes structured output, parallel tool calling, a Files API, prompt caching, and a built-in web_search tool.
Pricing and availability
Developer access is priced at US$1.25 per million input tokens and US$4.25 per million output tokens (approximately £1.00 and £3.40 respectively). New accounts receive US$20 in free credits. The public preview is US-only for now, with no European Union availability confirmed. Consumer access remains free in “Thinking” mode via the Meta AI app and meta.ai.
The Meta Model API is designed to be a drop-in replacement for existing workflows. It is OpenAI SDK-compatible — migration requires only changing the base URL to `https://api.meta.ai/v1` — and can accept Anthropic-format harnesses through the Messages API. Agent CLIs such as OpenCode can register Muse Spark 1.1 as a provider with a simple configuration change.
Benchmark positioning
Meta’s reported benchmarks position Muse Spark 1.1 as a tool-use and tool-augmented reasoning leader, while trailing established models on pure coding and multimodal tasks:
| Benchmark | Focus | Muse Spark 1.1 | Anthropic Opus 4.8 | GPT-5.5 (xhigh) | Gemini 3.1 Pro |
|———–|——-|:————–:|:——————:|:—————-:|:—————:|
| MCP Atlas | Scaled tool use | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench | Professional tool use | 54.7 | 48.4 | 38.3 | 15.9 |
| OSWorld-Verified | Computer use | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro | Repository-level coding | 61.5 | 69.2 | 58.6 | 54.2 |
| BabyVision | Visual reasoning | 76.3 | 81.2 | 83.6 | 51.5 |
The model leads on MCP Atlas and JobBench — benchmarks that measure how effectively a model can discover, orchestrate, and use external tools. It trails on SWE-Bench Pro (repository-level coding) and OSWorld (computer use), suggesting that Muse Spark 1.1 is primarily an orchestration model rather than a coding-accuracy leader.
Key architectural feature: self-managed context
One of the model’s more interesting capabilities is automatic context compaction. Rather than relying on the developer to manage the million-token window, the model actively remembers actions, retrieves earlier work, and compacts what it keeps. In a multi-agent setup, it can act as the lead agent — gathering context, planning, and delegating to parallel subagents — or as a subagent itself, adhering to tasks, using tools, and escalating when needed.
Meta demonstrated use cases including generating a Facebook Marketplace listing from a smartphone video (extracting photos, reasoning about the product, operating the browser to publish), and screenshot-driven debugging where the model traces failures and validates fixes.
Strategy shift
The release signals a strategic shift for Meta. Where previous model releases were open-weights and aimed at building ecosystem goodwill, Muse Spark 1.1 is a commercial product competing directly with Anthropic’s Opus and OpenAI’s GPT models in the developer API market. The closed, hosted, metered-per-token model means Meta is now collecting revenue rather than giving away weights — and the OpenAI SDK compatibility means it is making it as easy as possible for developers to switch.
Sources: Meta Superintelligence Labs releases Muse Spark 1.1 (MarkTechPost, July 9, 2026); Meta Model API documentation