Meta's new frontier model Muse Wonder Full joins the AI top tier.

Meta description: Meta Muse Wonder and Muse Wonder Full improve AI benchmark scores, especially agentic coding, while raising costs across tasks and tokens for users.

Tags: Meta Muse Wonder, AI benchmarks, agentic AI, coding benchmarks, AI pricing


The launch of Meta Muse Wonder, a frontier AI model offered in two variants, puts both models among the best AI models available. On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full reaches 56 points, while Muse Wonder records 52 points. The previous Muse Spark 1.3 recorded 48 points, so the new release moves the series from a mid-tier position into frontier competition. For the online advertising company, this matters because the gains are concentrated in agentic work, the area where Meta is pitching a personal AI agent that can access calendars, messages, and documents.

Background: A Mid-Tier Predecessor and a Weaker Open Llama Stretch

Muse Spark 1.3 had been the reference point for the Muse series, and its 48-point score placed it below the most prominent frontier systems. The new launch follows a weaker stretch for Meta with its open Llama models. That backdrop makes the benchmark movement more important: Muse Wonder Full places in the top three on six out of ten benchmarks overall, suggesting the company has caught up to the leading frontier models rather than posting a narrow result.

Meta also is selling features that go beyond raw scores. Active Memory lets the model remember key information across conversations and recall it when needed. Active Focus is the model's ability to stick with difficult tasks, try alternative approaches, and ask follow-up questions when something isn't clear. Both features support the company's personal AI agent pitch.

Development: Agentic Gains, Coding Strength, and Math Progress

The largest improvements show up in agentic evaluations, tests that measure how well an AI model can handle complex tasks on its own. On Terminal-Bench 4.0, performance jumped from 33 to 57 percent. Coding also rises sharply. Muse Wonder Full reaches 68 points on the Coding Agents Index, matching Claude Code with Sonnet 5.5. Muse Wonder scores 63, up from 54 for the predecessor.

Math performance changes in a similar way. On ProofBench, a private benchmark for mathematical proofs, Muse Wonder Full solves 94 out of 100 problems, up from 58 for its predecessor. That puts it just five to six points behind the top models, which score 99 to 100. On Terminal-Bench 4, Muse Wonder Full doubled its score while using 30 percent fewer output tokens. Performance on finance and legal tasks stayed at predecessor levels, dropping up to two points.

The Vals AI Index gives another view. According to Vals AI, Muse Wonder Full (Max) debuts at fifth place and beats all OpenAI models. Only current Anthropic models score higher. Muse Wonder lands at tenth.

Meta addresses a common criticism in its blog post, saying Muse models have been accused of being tuned for benchmarks but falling short in real-world use. The company says Muse Wonder Full does better than expected on unfamiliar tasks that weren't part of training optimization, and credits this to a stronger base model.

Latest Progress: Knowledge Gaps and a Steeper Price

Factual knowledge remains the clearest weak point. On the AA-Omniscience knowledge test, Muse Wonder Full manages only 30 points, and Muse Wonder gets 25. That is well behind Claude Opus 5.5 (46), GPT-6 Astra (43), and Gemini 4 Argon (42), which suggests the Wonder series is built on a smaller base model than its competitors.

Low hallucination rates partially offset the knowledge gap. Muse Wonder Full cuts its hallucination rate from 33 to 23 percent compared with the predecessor. That is the second-lowest among all models scoring at least 45 index points. Only Gemini 4 Argon, at 15 percent, hallucinates less often.

Better performance also comes with higher cost. According to Artificial Analysis, Muse Wonder runs $4.52 per task on the Intelligence Index, nearly three times more than Muse Spark 1.3, which cost $1.60 per task. Token pricing has climbed too, with $4 per million input tokens and $20 for output, up from $1.25 and $4.25 for the predecessor. Muse Wonder also uses 28 percent more input tokens per task, and those account for 76 percent of total costs.

Pricing still places Muse Wonder below Claude Fable 5.1 ($7.63) and Claude Opus 5.5 ($5.98), but above GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99), and GPT-6.1 Sol ($0.72). No token pricing is available yet for Muse Wonder Full. Both models do use about 10 percent fewer output tokens than their predecessor.

Current Stage: Strong Agent Tools, but Cost and Knowledge Remain Open Questions

Muse Wonder and Muse Wonder Full are no longer mid-tier on the Artificial Analysis Intelligence Index v4.3. Their agentic coding and math gains show why Meta is positioning them as frontier models. The open questions are narrower: whether Active Memory and Active Focus deliver dependable agent behavior outside benchmark conditions, whether the AA-Omniscience gap can be closed, and whether users will absorb higher input-token costs when output-token use falls by about 10 percent.

Meta says it trained the models at its in-house Prometheus data center at gigawatt scale. More details are in the model card. For now, the release marks a clear step forward in benchmark standing, while factual knowledge and pricing remain the most visible constraints.