Meta Enters the Developer API Market: Muse Spark 1.1 Ships With Native MCP, 1M Context, and a Drop-In SDK Swap

TL;DR: Meta's first commercial API model ships today — $1.25/$4.25 per M input/output tokens, native MCP and parallel tool calling, 1M-token context, and a base_url swap away from any existing OpenAI SDK integration.

Meta Superintelligence Labs launched Muse Spark 1.1 today, the first time Meta has charged for API access to one of its models. Until now Meta's models were open-weight and free; Muse Spark 1.1 is closed-weights and built specifically for agentic tool use and coding at scale.

Pricing in context

Model	Input	Output (per M tokens)
Muse Spark 1.1	$1.25	$4.25
GPT-5.6 Sol	$5.00	$30.00
Claude Opus 4.8	$5.00	$25.00

Output is 6× cheaper than GPT-5.6 Sol on sticker price. $20 in free credits for new API accounts; US-only preview at launch.

What makes it worth evaluating

• Native MCP: speaks the Model Context Protocol natively — your existing MCP servers connect without a wrapper layer
• Parallel tool calling: zero-shot generalization to new tools in structured output mode
• 1M-token context with self-managed compaction
• Computer use: direct desktop and browser control via screenshots
• Benchmarks: #1 on JobBench (professional tool use: 54.7 vs Opus 48.4) and MCP Atlas (88.1); trails on SWE-Bench Pro coding accuracy (61.5 vs Opus 69.2)

Drop it into your existing integration

The API is compatible with both OpenAI Chat Completions and Anthropic Messages formats — a base URL change is all you need:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_META_API_KEY",
    base_url="https://api.meta.ai/v1"
)

response = client.chat.completions.create(
    model="muse-spark-1.1",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize this PR diff and suggest edge cases..."}]
)

For MCP tool routing, pass your tool definitions in the same JSON schema format you use with Claude or GPT — Muse Spark 1.1 uses the same structure.

When to route to it

Route to Muse Spark 1.1 for high-volume agentic pipelines where output token cost is the bottleneck and coding accuracy is secondary (MCP tool calls, document processing, multi-step research). Keep Claude for precision coding and multi-file refactors where the SWE-Bench gap (61.5 vs 69.2) represents real risk.

Why it matters: A credible MCP-native competitor at 6× lower output cost forces every team building on GPT-5.x or Claude to reevaluate its routing strategy — today is a good day to benchmark your heaviest agentic workflows against Muse Spark's $4.25/M output.

Sources: Meta Muse Spark 1.1 Developer Guide (DigitalApplied) · Meta Starts Charging for AI With Muse Spark 1.1 (Bloomberg) · Muse Spark 1.1 pricing (AI Weekly)