Meta has launched its new Muse Wonder frontier AI models, which show significant performance gains in coding, math, and agentic tasks compared to the previous Muse Spark 1.3.
Key facts
- •Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index v4.3, while Muse Wonder scored 52.
- •The models feature 'Active Memory' for information recall and 'Active Focus' for handling complex tasks.
- •Muse Wonder Full solved 94 out of 100 mathematical problems on ProofBench, a significant increase from the predecessor's 58.
- •Hallucination rates for Muse Wonder Full dropped to 23 percent, down from 33 percent in the previous version.
- •Muse Wonder costs $4.52 per task, with input tokens priced at $4 per million and output tokens at $20 per million.
Meta has introduced two new frontier AI models, Muse Wonder and Muse Wonder Full. Both models demonstrate improved performance across several benchmarks, with Muse Wonder Full scoring 56 points on the Artificial Analysis Intelligence Index v4.3 and Muse Wonder scoring 52. These models feature new capabilities including 'Active Memory' and 'Active Focus,' designed to support personal AI agents that can interact with calendars, messages, and documents.
By the numbers
Performance and Capabilities
The new models show substantial improvements in agentic evaluations, with Muse Wonder Full achieving a 57 percent score on Terminal-Bench 4.0, up from 33 percent. In coding, Muse Wonder Full reached 68 points on the Coding Agents Index, tying with Claude Code. Mathematical performance also saw gains, with Muse Wonder Full solving 94 out of 100 problems on ProofBench. However, factual knowledge remains a challenge, with the models scoring lower than competitors like Claude Opus 5.5 and GPT-6 Astra on the AA-Omniscience test.
Features and Availability
Both models support a one-million-token context window and accept text, images, video, and audio inputs. Muse Wonder is currently available via Meta's API and Muse Code, while Muse Wonder Full is restricted to Muse Code users. Meta plans to integrate Muse Wonder into its app for paying subscribers. The company states the models were trained at its in-house Prometheus data center at gigawatt scale.
Pricing and Efficiency
According to Artificial Analysis, Muse Wonder costs $4.52 per task, compared to $1.60 for the predecessor, Muse Spark 1.3. Token pricing has increased to $4 per million input tokens and $20 for output tokens. Despite the cost increase, the models use approximately 10 percent fewer output tokens than their predecessor, and Muse Wonder Full uses 30 percent fewer output tokens on Terminal-Bench 4.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
