Meta Muse Spark 1.1 Developer Guide: First Paid Meta API for Agentic Tasks

Meta released Muse Spark 1.1 on July 9, 2026 and for the first time opened one of its in-house foundation models to outside developers through a new Meta Model API. This is Meta's first paid AI model - a multimodal reasoning model built for long, tool-heavy tasks that require planning and orchestration across apps and services.

Official Sources#

What Muse Spark 1.1 Actually Does#

Muse Spark 1.1 is a closed-source multimodal reasoning model optimized for personal agentic tasks. The key capabilities:

• 1 million token context window - large enough to process entire codebases in a single session
• Active context management - the model can compact context while preserving critical steps for later work
• Zero-shot tool generalization - works with native tools, MCP servers, and custom skills without fine-tuning
• Multimodal input - images, video, and PDF processing
• Computer use - can write scripts, navigate UIs, and orchestrate workflows across applications

Meta positions this against GPT-5.5 and Opus 4.8 on agentic evaluations, claiming top rankings on MedScribe, TaxEval, and Harvey's Legal Agent Bench while being "10x cheaper and twice as fast."

Pricing#

The pricing positions Muse Spark 1.1 as a budget option for high-volume agentic workloads. At these rates, a typical 50K input / 2K output agentic turn costs about $0.07 - roughly half of what you'd pay for Claude Sonnet 5 at intro pricing.

API Access#

The Meta Model API is currently in public preview for US-based developers. The API uses an OpenAI-compatible format, which means existing code using the OpenAI SDK can switch endpoints with minimal changes:

import OpenAI from 'openai';

const meta = new OpenAI({
  apiKey: process.env.META_API_KEY,
  baseURL: 'https://api.meta.ai/v1',
});

const response = await meta.chat.completions.create({
  model: 'muse-spark-1.1',
  messages: [
    { role: 'user', content: 'Analyze this codebase for security issues' }
  ],
  tools: [
    {
      type: 'function',
      function: {
        name: 'read_file',
        description: 'Read a file from the repository',
        parameters: {
          type: 'object',
          properties: {
            path: { type: 'string', description: 'File path' }
          },
          required: ['path']
        }
      }
    }
  ]
});

Key Features for Developers#

Tool Calling#

Muse Spark 1.1 supports structured tool calling with parallel execution:

const response = await meta.chat.completions.create({
  model: 'muse-spark-1.1',
  messages: [{ role: 'user', content: 'Find all TODO comments and create issues for them' }],
  tools: [readFileTool, listFilesTool, createIssueTool],
  parallel_tool_calls: true,
});

MCP Server Compatibility#

The model works with Model Context Protocol servers out of the box. If you're already using MCP with Claude Code or another MCP client, Muse Spark 1.1 can use the same server implementations without modification.

Multi-Agent Orchestration#

Muse Spark 1.1 can run as either a primary agent coordinating subagents or as a subagent itself. The model handles multi-turn agentic interactions with context compaction when approaching the 1M token limit.

Thinking Mode#

Available in the Meta AI app, Thinking mode shows the model's reasoning process before it produces a final response - similar to extended thinking in Claude models. API access to reasoning tokens is not yet documented.

How It Compares#

Pricing is the standout differentiator. At $1.25/$4.25 per million tokens, Muse Spark 1.1 is roughly 12x cheaper on input and 18x cheaper on output than Opus 4.8, while Meta claims competitive benchmark performance.

Early Partners#

Meta named three early API partners:

• Replit - integrated for agentic coding workflows
• Cline - using Muse Spark for their open-source coding agent
• Box - enterprise document processing pipelines

These integrations suggest Meta is targeting the same agentic coding and enterprise automation market that Anthropic and OpenAI dominate.

Limitations and Caveats#

US-only preview. The Meta Model API is currently limited to US-based developers. International availability isn't announced.

Closed source. Unlike Llama, Muse Spark 1.1 is proprietary. You can't self-host or inspect the weights.

No detailed benchmarks published. Meta claims competitive performance but hasn't released SWE-bench or other standardized coding benchmark scores. The comparison claims ("rivals GPT-5.5 and Opus 4.8") are marketing language until verified independently.

Preview status. Production guarantees and SLAs aren't documented. This is explicitly a preview, not GA.

When to Use Muse Spark 1.1#

Good fit:

• High-volume agentic workloads where cost matters more than bleeding-edge performance
• Tasks requiring very long context (full codebase analysis, long document processing)
• Teams already using OpenAI SDKs who want to test a cheaper alternative

Not a fit:

• Production workloads requiring SLAs (preview status)
• International teams (US-only)
• Tasks where you need published benchmark verification before committing

Getting Started#

1. Sign up at developer.meta.com with a US-based account
2. Navigate to the Model API section and create an API key
3. Claim your $20 free credits
4. Use the OpenAI-compatible endpoint at https://api.meta.ai/v1

FAQ#

Is Muse Spark 1.1 the same as Llama?#

No. Llama models are open-weights and can be self-hosted. Muse Spark 1.1 is a closed-source proprietary model only available through the Meta Model API.

Can I use Muse Spark 1.1 outside the US?#

Not currently. The public preview is limited to US-based developers. Meta hasn't announced international availability.

How does Muse Spark 1.1 handle tool calling?#

The API uses the same tool calling format as the OpenAI API, including parallel tool calls. Tools are defined as JSON schemas and the model returns structured tool call objects.

Is there a rate limit?#

Rate limits aren't documented in the preview announcement. Expect typical API rate limiting based on your account tier.

Does Muse Spark 1.1 support vision?#

Yes. The model accepts images, video, and PDFs as input. It can generate captions, analyze visual content, and produce code from visual designs.

How does the 1M token context compare to competitors?#

It's the largest publicly available context window from a major provider. Claude Opus 4.8 offers 200K tokens, GPT-5.5 offers 256K tokens. The 1M window is genuinely useful for full-codebase analysis without chunking.

What's the difference between Muse Spark and Meta AI?#

Meta AI is the consumer chat product (meta.ai). Muse Spark 1.1 is the underlying model now exposed through the developer API for programmatic access.

Can I use my existing OpenAI SDK code?#

Yes. The Meta Model API is OpenAI-compatible. Change the base URL and API key, and your existing code should work with minimal modifications.

Continue Reading#

• How We Patched 100+ PRs Across Our App Empire in One Day
• Running Fable 5 Agent Fleets in Production: The Operations Guide
• Running Fable 5 Agents on Vercel's eve Framework
• Meta Ships Muse Code and Muse Spark 1.2: A Terminal Agent With a 12x Cheaper Contributor Tier