On July 9, 2026, Meta Superintelligence Labs released Muse Spark 1.1, marking a significant performance jump in its proprietary model line. According to the Artificial Analysis Coding Index, the model achieved a score of 71, a 12-point increase over the 1.0 version released just three months prior. In the SciCode benchmark, Muse Spark 1.1 now sits at 58%, placing it third globally behind Claude Fable 5 at 60.2% and Gemini 3.1 Pro Preview at 58.9%. This rapid iteration cycle, coupled with an eight-point gain in the Overall Intelligence Index, signals that Meta is aggressively closing the capability gap with established frontier leaders.
The economic implications are immediate. Meta has priced Muse Spark 1.1 at $1.25 per million tokens for input and $4.25 for output, with a cached input rate of $0.15. This pricing structure undercuts current market leaders. Claude Sonnet 5 is currently priced at $2 for input and $10 for output under its introductory offer, which expires August 31. Even when compared to the most efficient tier of the GPT-5.6 series, Meta maintains a cost advantage: the $4.25 output price for Muse Spark 1.1 is lower than the $6 charged for GPT-5.6 Luna, the smallest and cheapest variant in the OpenAI lineup.
This release clarifies Meta’s strategic direction, which has moved beyond its identity as the primary steward of open-weight Llama models. The Meta Model API, which hosts Muse Spark 1.1, is a distinct product line. It is proprietary, closed-source, and cloud-only, offering no downloadable weights or self-hosting capabilities. The infrastructure powering Muse Spark is entirely new and explicitly not derived from the Llama architecture. There is no migration path between the two, forcing enterprise users to choose between the open-weight ecosystem and Meta’s new proprietary cloud service.
The development of this infrastructure is overseen by Alexandr Wang, the former CEO of Scale AI who now serves as Chief AI Officer at Meta and leads Meta Superintelligence Labs. The organizational structure reflects a deep commitment to this proprietary path: Meta invested $14.3 billion for approximately 49% of Scale AI as part of the deal that brought Wang to Meta in June 2025, according to multiple reports. By placing Wang at the helm of MSL, Meta has signaled that its future in frontier AI is tied to the same high-intensity data and engineering rigor that defined his previous tenure.
For agent builders, the field has effectively consolidated into three primary options: OpenAI’s GPT-5.6 series, Anthropic’s Claude, and now Meta’s Muse Spark 1.1. The availability of a third, high-performance option at a lower price point provides developers with greater leverage in managing operational costs for agentic coding tasks. As builders evaluate models based on a combination of capability, price, and API reliability, the entry of a lower-cost, high-capability provider forces a re-evaluation of existing vendor lock-in calculations.
The shift to a proprietary model represents a structural tension for Meta. While the company remains the dominant force in open-weight development, the Muse Spark 1.1 launch demonstrates that its most advanced capabilities are now reserved for a closed, cloud-based environment. This move mirrors the pricing compression dynamic previously established by the LongCat-2.0 model from Meituan, which proved that frontier-level capability could be delivered at commodity pricing. Meta is now applying this same pressure to the American hyperscaler tier.
The competitive dynamic is no longer defined by a binary choice between open-weight and proprietary models. It is increasingly defined by the ability of hyperscalers to deliver frontier performance at aggressive price points. By extending the commodity-pricing thesis to its own proprietary stack, Meta has introduced a new variable into the cost-benefit analysis for enterprise AI decision-makers. The result is a more crowded, price-sensitive market where the barrier to entry for high-end coding agents has been lowered considerably.