What is Muse Spark 1.1?
Meta Superintelligence Labs introduced a new multimodal reasoning model, Muse Spark 1.1, on July 9, 2026. As Meta states in its official blog post, the model is designed for agentic workflows: tool and computer usage across multiple applications, multi-agent orchestration, as well as coding tasks including bug fixes and code migrations in a corporate context. According to Meta, the context window accommodates up to one million tokens with active context management. Additionally, there is improved resilience against jailbreak and prompt injection attempts compared to the original Muse Spark version.
The model is initially available through a public preview of the new Meta Model API, which is designed to be compatible with existing OpenAI interfaces, as well as in the “Thinking” mode of the Meta AI app and on meta.ai.
A pricing model that undercuts the competition
The real break from Meta’s previous strategy lies in the business model: it is the first paid developer API for one of Meta’s own frontier models. According to reports from Fortune and Bloomberg, Meta charges $1.25 per million input tokens and $4.25 per million output tokens; cached input tokens cost $0.15 per million, and a web search connection with citations costs $2.50 per 1,000 queries. New developers receive a $20 credit to get started.
This positions Muse Spark 1.1 significantly below the prices that Anthropic, OpenAI, and other providers charge for their respective top models – according to an analysis by Techzine Global, the model is priced between smaller models like GPT-5 mini or Haiku 4.5 and medium-sized models like Claude Sonnet 4.6, while offering performance metrics that, according to Meta, reach those of significantly more expensive top models in certain categories. Notably, the timing coincides: just a day earlier, xAI had released a model with a similarly aggressive pricing strategy, Grok 4.5 – both providers are explicitly positioning themselves as affordable alternatives to established agent models.
Benchmarks: Strong in agents, weaker in pure coding
According to Meta’s own metrics, Muse Spark 1.1 achieves top scores in several agent and tool use benchmarks: 88.1 points in the MCP Atlas test for scaled tool handling and 54.7 points in the JobBench test for professional tool usage – both ahead of Claude Opus 4.8 and GPT-5.5. The model also scores 62.1 points in the reasoning test Humanity’s Last Exam, placing it ahead of both comparison models.
In pure coding benchmarks like SWE-Bench Pro, however, Muse Spark 1.1 falls behind Claude Opus 4.8, as well as in multimodal tasks. Fortune reports that the model also lags behind Anthropic’s Mythos 5 and Fable 5, as well as OpenAI’s GPT-5.6 in certain coding metrics. Important for context: all mentioned values come from Meta’s own test series and have not yet been independently verified. The independent analysis portal Artificial Analysis had not conducted its own measurements for version 1.1 at the time of publication, only for the original Muse Spark version released in April 2026.
Strategic shift: Meta departs from the open Llama approach
The re-release confirms a course change that Meta initiated in April 2026 with the first Muse Spark version: the departure from the open Llama model approach in favor of closed models. According to Fortune’s reporting, CEO Mark Zuckerberg justified the focus by stating the aim to offer powerful agentic and multimodal models at the lowest possible costs. Responsible for the strategy is Alexandr Wang, Meta’s first Chief AI Officer since the reorganization to Meta Superintelligence Labs; Meta had previously acquired a non-voting stake of 49 percent in Wang’s former company Scale AI for $14.3 billion in 2025.
Classification
Muse Spark 1.1 is primarily a model for agent tasks with a corresponding business model, not an attempt to outperform the competition in every category. CNBC classifies the move as a direct push by Meta into the AI coding and agent tool market dominated by Anthropic and OpenAI. Whether the aggressive pricing structure will actually draw developers away from established providers remains to be seen, pending independent tests and a launch in the EU – both are still outstanding.
