Meta Muse Spark 1.1: Meta's First Paid AI Model

Meta launched Muse Spark 1.1 on July 9, 2026 — its first paid frontier AI model with a public developer API. Priced at $1.25 per million input tokens, it targets agentic coding, tool use, and computer-use workflows with a 1-million-token context window, directly competing with OpenAI and Anthropic's developer ecosystems.

I verified the pricing, benchmark claims, and launch timeline against Meta's official Muse Spark 1.1 announcement, the Artificial Analysis model page, and coverage from Reuters. The figures below are sourced from Meta's own benchmark table and independent trackers, not promotional copy.

Key Takeaways

Muse Spark 1.1 is Meta's first paid frontier model — a strategic shift from its open-weights playbook
Pricing: $1.25/1M input tokens, $4.25/1M output — competitive with GPT-5.5 and Claude Opus 4.8
1M token context window with OpenAI-compatible API, $20 free credits for US developers
Leads on agentic benchmarks (MCP Atlas 88.1, Finance Agent v2 57.2) but trails on coding (SWE-Bench Pro 61.5 vs Opus 4.8's 69.2)
Available in public preview for US developers via Meta Model API

What Is Muse Spark 1.1?

Muse Spark 1.1 is Meta's upgraded multimodal reasoning model from Meta Superintelligence Labs, first introduced in April 2026. The 1.1 release adds developer API access, a 1-million-token context window, and improved performance on agentic tasks — tool use, computer use, multi-step coding, and script generation. It is available through the new Meta Model API with OpenAI-compatible endpoints.

This is Meta's first paid frontier model, marking a clear departure from the open-weights strategy that defined Llama and earlier releases. Rather than distributing weights for self-hosting, Meta is now charging per-token for API access, directly competing with OpenAI's GPT-5.5/5.6 and Anthropic's Claude Opus 4.8 and Fable 5.

Unlike OpenAI's three-tier GPT-5.6 family which segments by capability and price, Muse Spark 1.1 ships as a single model with a straightforward pricing tier. The approach is simpler but less flexible — developers who need a cheaper option for simpler tasks have to look elsewhere.

Pricing and API Access

Muse Spark 1.1 costs $1.25 per million input tokens and $4.25 per million output tokens. Artificial Analysis estimates a blended rate of $0.78 per million tokens assuming a 7:2:1 cache-to-input-to-output ratio. US developers get $20 in free credits to start.

For context, this places Muse Spark 1.1 below GPT-5.5's pricing ($2.50/$10 per 1M tokens for the standard tier) and roughly on par with Claude Opus 4.8. Early reporting from 36Kr claims Meta is positioning it as "10 times cheaper than Claude Fable 5," though this figure should be treated as secondary reporting until Anthropic publishes official Fable 5 API pricing.

Benchmark Performance — Where It Leads and Where It Lags

Meta's official benchmark table compares Muse Spark 1.1 against the original Muse Spark, Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 across multiple categories.

Where it leads:

MCP Atlas (agentic tool use): 88.1 — highest among compared models
JobBench (real-world task completion): 54.7
Humanity's Last Exam with tools: 62.1
Finance Agent v2: 57.2

Where it lags:

SWE-Bench Pro (coding): 61.5 vs Claude Opus 4.8's 69.2
Terminal-Bench 2.1 / DeepSWE 1.1 / BabyVision: GPT-5.5 leads across all three

The pattern is clear: Muse Spark 1.1 excels at agentic orchestration and tool-use workflows but falls short on pure coding benchmarks. For developers building multi-agent systems or automation pipelines, it may be the better choice. For traditional code generation tasks, Claude Opus 4.8 or GPT-5.5 may still be preferable.

This aligns with our analysis of Anthropic's Claude models — different models optimize for different workloads, and the right choice depends on whether your bottleneck is coding accuracy or agentic reliability.

Strategic Shift: Meta's Open-Weight Era Ends

Muse Spark 1.1's paid API model represents a fundamental strategy change for Meta. After years of releasing Llama models as open weights for self-hosting, Meta is now joining the paid API race. The announcement mentions partners like Replit, Cline, Box, and OpenClaw Foundation — all building on the Meta Model API rather than self-hosted deployments.

For developers, this means one fewer open-source alternative in the frontier model space. If you've been relying on Meta's open models to avoid vendor lock-in, Muse Spark 1.1 won't fill that role. The company has not announced whether a smaller open variant is in development.

How Muse Spark 1.1 Fits Into Your AI Stack

Best for: Agentic automation, tool-use pipelines, computer-use agents, long-context tasks (1M tokens)
Not ideal for: Pure code generation, low-cost bulk inference, offline/self-hosted deployments
API compatibility: OpenAI-compatible — drop-in replacement for existing GPT/Claude integrations
Availability: US developers in public preview; global timeline not yet announced

Frequently Asked Questions

How much does Muse Spark 1.1 cost?

$1.25 per million input tokens and $4.25 per million output tokens. US developers receive $20 in free credits for the public preview.

Is Muse Spark 1.1 open source?

No. Unlike Meta's previous Llama models, Muse Spark 1.1 is a closed, paid API model. Meta has not announced an open-weight version.

How does Muse Spark 1.1 compare to GPT-5.6?

Muse Spark 1.1 leads on agentic benchmarks (MCP Atlas, Finance Agent) while GPT-5.5 leads on coding benchmarks. Pricing is similar per token, but GPT-5.6 offers tiered options (Sol, Terra, Luna) at different price points.

Can I use Muse Spark 1.1 outside the US?

The public preview is currently limited to US developers. Meta has not announced a timeline for global availability.

Sources