Meta upgrades Muse Spark to version 1.1 with improved reasoning and lower hallucination rates

1. Performance Gains

Meta has released Muse Spark 1.1, which achieved a score of 51 on the Artificial Analysis Intelligence Index. This represents an 8-point improvement over the previous version, Muse Spark 1.0, within a three-month period. The model’s performance gains are primarily attributed to advancements in scientific reasoning, coding, and knowledge-based tasks. Notable improvements include a 12-point increase in the Coding Index and a significant rise in the AA-Omniscience score, which climbed from 4 to 18. This specific gain in omniscience was driven by a 35-point reduction in the hallucination rate, as the model increasingly opted to abstain from answering rather than providing inaccurate information.

2. Efficiency and Cost

Muse Spark 1.1 is positioned as a cost-effective and token-efficient option compared to peer models with similar intelligence scores. During testing, the model utilized 94 million output tokens, which is lower than the consumption rates of comparable models like GLM-5.2, GPT-5.4, and GPT-5.6 Luna. Based on Meta’s pricing of $1.25 per 1 million input tokens and $4.25 per 1 million output tokens, the estimated cost per Intelligence Index task is approximately $0.26. This pricing structure makes it one of the more affordable models among those currently performing at the 51-point intelligence level.

3. Technical Specifications

The updated model features a context window of 1 million tokens, a significant increase from the 262,000-token capacity of its predecessor. In terms of speed, the model delivers a median output of approximately 114 tokens per second on Meta’s first-party API, with a time-to-first-token of about 21 seconds. Muse Spark 1.1 is currently available through Meta’s first-party API, and the company also offers discounted pricing for cache hits at $0.15 per 1 million tokens.