Muse Spark 1.1 has credible high-end capability evidence, but not evidence of leadership. Meta reports 77.4% SWE-Bench Verified under a 15-attempt methodology; LiveCodeBench tracking reports 69.7% (15th/72), while DeepSWE reports 53% ±3% (11th/18). This supports coding/agentic performance around DeepSeek-V3.2 and below Kimi K2.5, GPT-5.5, GPT-5.6 Sol, and Claude Opus 4.8. Meta’s own evaluation says it trails Opus 4.8 and/or GPT-5.5 on Terminal-Bench 2.1 and SWE-Bench Pro. Multimodal and API breadth are relative strengths: official material documents native multimodal reasoning, tool/function calling, and a managed 1M-token context. Its 51 Artificial Analysis Intelligence Index and 1505±7 Arena hard-prompts result (14th, preliminary; 8,670 votes) indicate solid general reasoning and preference, not frontier placement. The $1.25/M input and $4.25/M output pricing is materially more economical than premium frontier anchors, while AA’s 127.3 tokens/s and 2.47s TTFT support a top-tier speed score. Public preview status tempers availability and adoption. Developer experience is above GLM-4.6 and DeepSeek-V3.2 due to Meta-hosted API access and agentic affordances, but below GPT-5.5’s more mature tooling ecosystem. Risk/evidence quality is relatively strong: Meta provides evaluation, safety, and preparedness reporting, yet key performance claims remain substantially vendor-reported and several independent benchmark families provide limited or no directly comparable detail. Pricing is independently listed, but official pricing and commercial terms were not supplied.
