Meta Enters AI Top Tier with Muse Wonder and Muse Wonder Full Frontier Models

Meta has officially released two variants of its latest frontier AI model: Muse Wonder and Muse Wonder Full. The new releases signal a major technical leap for the tech giant, lifting its models directly into the top tier of competitive benchmarks after a weaker period with its open Llama systems.

On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scored 56 points, while Muse Wonder registered 52 points. Both represent a notable jump over their predecessor, Muse Spark 1.3, which reached 48 points. Across the board, Muse Wonder Full secured top-three placements in six out of ten evaluated benchmarks, while debuting at fifth place on the Vals AI Index—beating all OpenAI models and trailing only Anthropic's flagship systems.

Leaps in Agentic Coding and Problem-Solving

The most substantial advances in the Wonder generation appear in agentic evaluations, which measure a model's capacity to execute multi-step workflows without continuous human prompting. On Terminal-Bench 4.0, task performance increased from 33 percent to 57 percent, while simultaneously requiring 30 percent fewer output tokens.

Software development saw an equally sharp boost. On the Coding Agents Index, Muse Wonder Full scored 68 points, tying Claude Code with Sonnet 5.5. Muse Wonder reached 63 points, up from 54 points achieved by Muse Spark 1.3. In mathematics, Muse Wonder Full solved 94 out of 100 problems on the private proof benchmark ProofBench, up from 58 problems solved by the previous version.

Meta attributes these gains to two primary architectural behaviors:

• Active Memory: Allows the model to retain and recall key information across multi-turn sessions.
• Active Focus: Enables persistent reasoning through challenging instructions, prompting the model to explore alternate strategies or ask clarifying questions when encountering ambiguous context.

Both models feature a one-million-token context window and support multimodal input spanning text, images, video, and audio. Meta trained the series at gigawatt scale at its dedicated Prometheus data center.

Knowledge Limitations and Hallucination Rates

While Muse Wonder demonstrates elite performance in coding and autonomous execution, factual retrieval remains a noticeable vulnerability. On the AA-Omniscience test, Muse Wonder Full scored 30 points and Muse Wonder scored 25 points. This places Meta far behind competing frontier models, including Claude Opus 5.5 (46 points), GPT-6 Astra (43 points), and Gemini 4 Argon (42 points), suggesting that the Wonder line relies on a comparatively compact base architecture.

Performance on finance and legal tasks also remained flat, showing minor declines of up to two points compared to the previous release.

However, Meta offset retrieval gaps with improved reliability. Muse Wonder Full lowered its hallucination rate from 33 percent to 23 percent. Across all models scoring at least 45 points on the Intelligence Index, only Gemini 4 Argon (15 percent) recorded fewer hallucinations. Meta also noted that the model maintained high resilience on unfamiliar tasks outside its specific training optimizations.

Escalating API Costs and Deployment

The performance gains come alongside a significant cost increase. Artificial Analysis reports that running Muse Wonder costs $4.52 per task on the Intelligence Index, compared to $1.60 for Muse Spark 1.3.

Per-token rates have shifted to $4 per million input tokens and $20 per million output tokens (up from $1.25 input and $4.25 output on the prior model). Input tokens account for 76 percent of the total cost due to a 28 percent increase in input consumption per task, though overall output generation was roughly 10 percent leaner. This pricing places Muse Wonder above GPT-6 Astra ($3.26) and Gemini 4 Argon ($1.99), but below Claude Opus 5.5 ($5.98) and Claude Fable 5.1 ($7.63). Token pricing for Muse Wonder Full has not yet been announced.

Muse Wonder is currently available via Meta's API and Muse Code, with plans to roll it out to paying subscribers of the Muse consumer app before releasing specialized variants. Muse Wonder Full remains exclusively accessible to Muse Code users.