Meta launched Muse Spark 1.1 and the new Meta Model API – its first proprietary, paid AI coding agent built to compete directly with Anthropic’s Claude and OpenAI’s GPT families.
This is a major strategic pivot for Meta. After Llama 4 drew a lukewarm developer response, Meta formed Meta Superintelligence Labs led by Scale AI founder Alexandr Wang and repositioned from open-source champion to closed-model provider. CEO Mark Zuckerberg broke a 3-year silence on X to call it “a strong agentic and coding model at a very low price” that is “strongest at agentic performance, tool use, and computer use”.
Muse Spark 1.1 is not just a chatbot. It is designed for large agentic workloads, bug fixing, and large code migrations – enterprise automation tasks where Anthropic and OpenAI currently dominate. Its edge is efficiency + price.
Key Details
1. What it is
- Name: Muse Spark 1.1 (original Muse Spark codenamed “Avocado” launched April 2026). Released April 8, 2026; 1.1 on July 9, 2026.
- Builder: First model from Meta Superintelligence Labs.
- Type: Natively multimodal reasoning model – processes voice, text, and visuals simultaneously, built “from the ground up to integrate visual information across domains and tools”.
- Positioning: Meta describes it as “the first step on our scaling ladder” toward personal superintelligence.
2. Technical Architecture
- Context Window: 1 million tokens, up from ∼262K on original Muse Spark. Critical for repository-level coding.
- Agentic Design: Multistep reasoning, digital workflow management, deploying new features in enterprise systems. Delivers “exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services”.
- Contemplating Mode: Uses multiple parallel agents to solve hard problems, achieving 58% on Humanity’s Last Exam and 38% on FrontierScience Research in internal testing.
- Efficiency Tech: “Thought compression” – after initial longer thinking, length penalty compresses reasoning to use far fewer tokens. Burns 58M output tokens for full Intelligence Index run vs 157M for Claude Opus 4.6 and 120M for GPT-5.4.
- Speed: ∼114 tokens/sec median on Meta’s API, ∼21 sec time to first token.
3. Benchmarks & How It Stacks Up
- Intelligence Index: 52 points, top 5 globally, trailing only Gemini 3.1 Pro Preview, GPT-5.4, Claude Opus 4.6. Llama 4 Maverick had 18.
- Coding Benchmarks (mixed):
- SWE-Bench Pro: 61.5% vs Claude Opus 4.8 at 69.2%
- Terminal-Bench 2.1: 80.0% vs GPT-5.5 at 83.4%
- DeepSWE 1.1: 53.3% vs GPT-5.5 at 67.0%
- Tool-Use Strengths: Leads on agentic tool-use: MCP Atlas 88.1, JobBench 54.7
- Cost Efficiency: $0.26 per task vs $0.37 for GLM-5.2 and $0.89 for GPT-5.4, using only 94M output tokens vs 141M for GLM-5.2
Interpretation: Not yet frontier on raw coding reasoning, but best-in-class on tool use and cost – the enterprise sweet spot.
4. Business Model & Pricing
This is Meta’s first paid model. No self-hosting, not open-source – proprietary and served only from Meta’s properties.
- Pricing: $1.25 per million input tokens, $4.25 per million output tokens. In line with Anthropic Claude Haiku 4.5 and OpenAI GPT-5.6 Luna, but slightly above them. Wang called it “very aggressive and attractive”.
- API: Meta Model API in public preview, OpenAI-compatible, with $20 free credits on signup.
- Availability: Live on Meta AI web portal and mobile app, with plans to integrate into WhatsApp, Instagram, and Meta smart glasses.
5. Why It Matters in the Competitive Race
- For Meta: Signals end of “open-source at all costs” era. From Llama 2-based Code Llama in 2023 to a closed, monetizable developer platform. It also validates the $B+ bet on Wang and Superintelligence Labs.
- For Anthropic / OpenAI: Pressure on price. Meta undercuts by ∼4x on input/output vs Opus and GPT-5 class, targeting the high-volume agentic coding market where token costs compound fast.
- For Enterprises: Ideal for large codebase migrations, automated bug fixes, and orchestration across external apps – not just autocomplete, but autonomous agent workflows.
Limitations: Meta admits gaps remain in coding and “long-horizon agentic systems”. It still trails Opus 4.8 and GPT-5.5 on hard coding benchmarks.
