Meta's Muse Spark 1.2 Boosts Agentic Capabilities, Ties for 3rd Among US Labs
ArtificialAnlys · x · 2026-08-06
Meta has released Muse Spark 1.2, its third model upgrade in four months. Scoring 54 on the Artificial Analysis Intelligence Index, it significantly improves agentic knowledge work capabilities, tying Meta with SpaceXAI for 3rd place among US labs.
Key Evaluation Details:
- Major Agentic Gains: GDPval-AA v2 Elo surged 260 points to 1631 (ranking #5), with minor gains in Terminal-Bench 2.1 and Tau3-Bench Banking.
- High Cost-Efficiency: At $0.40 per Intelligence Index task, it is among the cheapest in its intelligence cluster, behind only Grok 4.5 and GPT-5.6 Sol.
- Lower Hallucination via Abstention: The hallucination rate dropped 10 points to 28% by increasing the abstention rate (attempt rate fell to 67%), though overall accuracy slightly slipped to 38%.
- Pricing & Context: Maintains a 1M token context window and unchanged pricing ($1.25/$4.25 per 1M input/output tokens).
Related event: Meta Releases Muse Spark 1.2 with Improved Agent Capabilities(3 posts)→
More from Models
- Hands-on: Meta's Muse Model Delivers Instant Replies on WhatsApp — DevDminGod · 2026-08-06
- Trending on HF: 10M+ Multilingual LLM Distillation Dataset — r0b0tlab · 2026-08-06
- Reddit user predicts upgrades in Gemini 3.5 Pro — Last_Conclusion_8984 · 2026-08-06
- KOL Slams Anthropic Models as Overpriced, Warns of Losing Market Edge — robleclerc · 2026-08-06
- ARC-AGI 3 benchmark fallen; Ilya's SSI model crushes it harness-free — iruletheworldmo · 2026-08-06
- Muse Spark 1.2 Scores 54 on Intelligence Index with Competitive Pricing — ArtificialAnlys · 2026-08-06