Meta has officially re-entered the AI arms race with the launch of Muse Spark 1.1, a multimodal model designed specifically for agentic tasks and coding. The announcement, made on July 9, 2026, marks a significant strategic shift for the company as it moves beyond its traditional open-source approach to offer a paid API for developers. With a price point roughly 25% that of top-tier models from OpenAI and Anthropic, Meta is betting that affordability, combined with competitive performance, can carve out a substantial share of the enterprise AI market. The release even prompted Meta CEO Mark Zuckerberg to post on X for the first time in three years, signaling the importance of this product.

A Strategic Leap into Agentic AI and Coding

Muse Spark 1.1 is the second product from Meta’s Superintelligence Labs (MSL), led by former Scale AI CEO Alexandr Wang. The model is built to orchestrate multiagent systems, where a primary agent creates a plan and delegates tasks to parallel subagents. This architecture allows the model to handle complex, multi-step projects with minimal human intervention, a capability increasingly sought after by enterprises. Meta claims the model can manage a 1-million-token context window, actively remembering actions and retrieving information from earlier work to maintain coherence across long sessions. In internal testing, the model reportedly outperformed Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5 on four agentic benchmarks, including the demanding Humanity’s Last Exam.

Context Window and Capabilities | Feature | Specification | |—|—| | Context Window | 1 million tokens | | Input Modalities | Text, images, video, audio, PDFs | | Core Strengths | Multiagent orchestration, computer use, coding, tool use |

Performance and Benchmarks: A Mixed but Promising Picture

While Meta touts Muse Spark 1.1 as a leader in agentic performance, the benchmark results reveal a more nuanced story. The model scored a 72 on “vibe coding” benchmarks, a massive leap from the original Muse Spark’s score of under 20. However, it lagged behind competitors in some pure coding and multimodal tasks, though the gap was significantly smaller than with the previous version. Wang described the model as “the strongest model for agentic and coding work yet” from Meta, emphasizing that coding ability is a foundational component of overall agentic capability. The model also excels in computer use workflows, navigating unfamiliar interfaces and automating tasks across multiple applications, a feature that aligns with the industry trend toward autonomous digital assistants.

Key Benchmark Performance | Benchmark | Muse Spark 1.1 | Muse Spark (Original) | Competitors (e.g., Opus 4.8, GPT-5.5) | |—|—|—|—| | Agentic Tasks (Humanity’s Last Exam) | Outperformed | N/A | Outperformed | | Vibe Coding | 72 | <20 | Comparable | | Multimodal Reasoning | Competitive, slight lag | Significant lag | Leading |

Pricing Strategy: The Cheapest Frontier Model on the Market

Meta’s most disruptive move is its aggressive pricing. The company will charge $1.25 per million input tokens and $4.25 per million output tokens for Muse Spark 1.1. This is roughly one-quarter the cost of equivalent models from OpenAI and Anthropic. While this price is higher than OpenAI’s GPT-5 mini and Anthropic’s Claude Haiku 4.5, it sits at a similar level to GPT-5.6 Luna and is significantly cheaper than Anthropic’s Claude Sonnet 4.6. Each new API account will receive $20 in free credits to test the model. Zuckerberg framed this as a democratization effort, contrasting Meta’s approach with what he described as competitors “keeping a model for themselves.” This pricing strategy is crucial for Meta, which has committed hundreds of billions of dollars to AI infrastructure and is seeking new revenue streams to justify the investment.

Pricing Comparison (Per Million Tokens) | Model | Input Cost (USD) | Output Cost (USD) | |—|—|—| | Muse Spark 1.1 | $1.25 | $4.25 | | GPT-5 mini (OpenAI) | < $1.25 | < $4.25 | | Claude Haiku 4.5 (Anthropic) | < $1.25 | < $4.25 | | GPT-5.6 Luna (OpenAI) | ~$1.25 | ~$4.25 | | Claude Sonnet 4.6 (Anthropic) | > $4.25 | > $4.25 |

Multimodal Capabilities and Safety

Muse Spark 1.1 is a truly multimodal model, capable of processing text, images, video, audio, and PDFs. It can “see and hear” to complete tasks, interpreting what is on a screen and taking action. This includes generating “ultra-descriptive” captions from visual and audio inputs. On the safety front, Meta has conducted extensive evaluations following its Advanced AI Scaling Framework, claiming the model shows strong resistance to jailbreaks, prompt injection, and other common attacks.

The Future: Watermelon and the Open-Source Dilemma

Meta is already looking ahead. Wang revealed that MSL is training a more powerful model codenamed “Watermelon,” which has reportedly caught up to OpenAI’s GPT-5.5 on key benchmarks. Notably, while Meta is charging for Muse Spark 1.1, the company remains committed to open source. Wang stated that the MSL team is developing a variant of Muse Spark that will be released as open source, though no timeline was given. This dual strategy—offering a paid, high-performance model while maintaining an open-source presence—could allow Meta to compete on multiple fronts. The new model is also expected to replace some existing Llama models powering Meta’s own products, including WhatsApp, Instagram, and Facebook chatbots.