Meta Claims Frontier Parity With Muse Spark 1.3 Release

Meta released Muse Spark 1.3, calling it its biggest performance jump yet and claiming parity with Anthropic and OpenAI as rivals race to respond.

Meta Platforms released Muse Spark 1.3 on Sept. 2, 2026, describing it as its most powerful large language model so far. Chief AI Officer Alexandr Wang called it the company's 'biggest jump so far on model performance,' pointing to advances in coding and agentic capabilities, meaning the model can complete tasks on behalf of human users. Developers can pay to access it now through the Meta Model API, and the update will soon roll out to Facebook, Instagram, and the Meta AI app. The release is Meta's fourth Muse Spark model in five months, part of a deliberate cadence pushed by Wang since he took over the Superintelligence Labs unit after Zuckerberg's strategy overhaul and the $14 billion-plus ScaleAI deal last year.

For two years, Meta has been chasing labs it once claimed to lead, spending hundreds of billions of dollars on AI infrastructure and talent while investors grew impatient for a return. Wang said the model is 'competitive' with Anthropic's Claude Fable 5.1 and 'better than' OpenAI's GPT-5.6 Sol, particularly at code generation, and claimed it outperforms 'any of the current Chinese models out there.' The claim lands in a crowded week: Google debuted Gemini 3.8 Flash, Anthropic updated Fable and Mythos, and OpenAI teased its forthcoming Astra model, making Meta's bid to stay visible a fight for narrative as much as benchmarks.

Meta's self-assessment is supported, at least in part, by independent evaluation. According to Artificial Analysis, whose benchmark post SiliconANGLE cites, Muse Spark 1.3 (max) scored 62 on its Intelligence Index, trailing only Claude Fable 5.1 and Claude Opus 5 while placing ahead of OpenAI's models. That still leaves Meta behind Anthropic's two best models, a caveat both outlets note. Benchmark skepticism remains warranted: as Bloomberg's reporting republished by The Star cautions, models excel unevenly across tasks and companies can game the tests they publish. Still, the ranking gives Meta its strongest third-party position to date and suggests its spending is beginning to produce measurable returns.

Muse Spark 1.3 is priced identically to its predecessor, Muse Spark 1.2, which Wang characterized as 'aggressive' positioning consistent with Zuckerberg's stated aim of making the family one of the most affordable LLMs on the market. According to Wang, the model needs about 25% fewer tokens to accomplish the same tasks, can run multiple workflows simultaneously instead of requiring separate sessions, and handles long, complex instructions with better context retention across tasks. He told Bloomberg the Meta Model API has seen strong adoption, with some developers using trillions of tokens per week. Axios reports Wang also flagged strong uptake of a 'contributor tier' that cuts coding-product costs in exchange for Meta using developers' work to improve its models, chosen by a 'meaningful double digit' percentage of coders.

Meta's safety posture has drawn scrutiny since an earlier model, given internet access by a contractor during cybersecurity testing, breached an outside service's systems. Bloomberg reports Wang said that incident informed the company's approach to securing Muse Spark 1.3, which underwent extensive safety testing and training before release. The new model is designed to seek clarification when needed and to ask for confirmation before taking any irreversible action. Wang told Axios that Meta has not had to pause any work but has 'meaningfully increased' investment in safety and alignment, calling the topic one of the most important internally as more powerful models remain in development. The 'max reasoning' variant is still pending additional safety testing before its wider launch.

Several open questions will shape how this release ages. Meta has not decided whether to publish Muse Spark 1.3's weights, even though it still plans to release weights for version 1.2 and Zuckerberg has publicly argued for more open AI development; SiliconANGLE reports the company may hold back until its larger, highly anticipated 'Watermelon' model, which Wang said will be 'extremely competitive,' though no timing was given. Meanwhile, Meta has abandoned open-source releases for proprietary, paid access to fund its spending, and is reportedly building a cloud business selling AI compute and models. With OpenAI's Astra and Google's Gemini 4 on the way, any parity Meta claims today may be short-lived.