Meta's Muse Spark 1.1 Gains 8 Points on Intelligence Index, Matches Frontier Competitors

Summary

Meta has released Muse Spark 1.1, a significant update to its LLM that scores 51 on the Artificial Analysis Intelligence Index, up 8 points from Muse Spark 1.0's score of 43 just three months earlier. The new model effectively ties with GPT-5.4, GLM-5.2, and GPT-5.6 Luna while remaining three points behind Grok 4.5 and trailing frontier leaders Claude Fable 5 (60), GPT-5.6 Sol (59), and Claude Opus 4.8 (56).

Performance improvements are concentrated in Scientific Reasoning, Coding, and Knowledge work. Notably, Muse Spark 1.1 achieves 58% on SciCode, ranking third across all benchmarked models, and reaches 45% on Humanity's Last Exam—within one point of Claude Opus 4.8 despite being five points lower on the overall Intelligence Index. The model shows substantial gains in agentic knowledge work with a 232 Elo improvement on GDPval-AA v2.

The update brings significant technical enhancements alongside competitive economics. Muse Spark 1.1 expands context window to 1M tokens (up from 262k), prices at $1.25/$4.25 per 1M input/output tokens with cache hits at $0.15, and achieves estimated costs of ~$0.26 per Intelligence Index task. Most notably, it uses just 94M output tokens to run the Intelligence Index—fewer than competing models at similar performance levels—making it the most token-efficient among models scoring 51.

The gains were driven primarily by improved abstention patterns rather than accuracy alone. Muse Spark 1.1's hallucination rate fell 35 percentage points to 38% while accuracy remained relatively flat, demonstrating Meta's focus on reliability and reduced false confidence in model outputs.

• Context window expands to 1M tokens with competitive pricing of $1.25/$4.25 per 1M tokens, plus $0.15 for cache hits
• Improvements driven by reduced hallucination (35-point drop to 38%) and improved abstention rather than raw accuracy gains