Meta Releases Muse Spark 1.3: 75.4% DeepSWE Score and $0.10 Contributor Pricing
Meta Superintelligence Labs officially deployed Muse Spark 1.3 across the Meta Model API, Muse Code, and partner platforms including OpenRouter on September 3, 2026, marking the company's third major frontier update in five months following Muse Spark 1.2 and Muse Glimmer.
Engineered specifically for high-autonomy software engineering and long-horizon tool orchestration, Muse Spark 1.3 operates inside a 1,048,576-token context window. According to technical documentation confirmed by Meta Chief AI Officer Alexandr Wang in interviews with Bloomberg and Axios, the model completes complex development tasks with approximately 20% fewer tool calls and 25% fewer generated tokens compared to Muse Spark 1.2.
On standardized benchmarks reported by DataCamp and Flowtivity, Muse Spark 1.3 achieved a leading 75.4% on DeepSWE 1.1 (surpassing Claude Opus 5's 74.0% and [GPT-5.6 Sol](/models/gpt-5-6-sol)'s 73.0%), tied GPT-5.6 Sol at 88.8% on Terminal-Bench 2.1, and posted 98.5% on multi-needle retrieval across 512K token contexts. Independent evaluation by Artificial Analysis recorded an Intelligence Index score of 61 on the xhigh variant at an inference speed of 182 tokens per second.
Meta is offering two primary commercial tiers: a heavily subsidized "Contributor" tier priced at $0.10 per million input tokens and $0.20 per million output tokens for developers opting into product-improvement training, and a standard enterprise "xhigh" tier at $1.25 input and $4.25 output per million tokens with an 88% prompt cache discount.
