Meta's new frontier model Muse Wonder Full breaks into the AI top tier

Meta has launched Muse Wonder and Muse Wonder Full, two variants of a new frontier model family that the company says lift it back among the strongest AI systems available. The release was accompanied by official benchmark results and a model card, and Artificial Analysis has published independent index scores.
The headline numbers are on the Artificial Analysis Intelligence Index v4.3. Muse Wonder Full reaches 56 points and Muse Wonder 52, against 48 for the predecessor Muse Spark 1.3. That is a meaningful climb from what the source describes as a mid-tier position. Muse Wonder Full places in the top three on six of ten benchmarks, and on the Vals AI Index it debuts fifth in its Max configuration, ahead of every OpenAI model listed by Vals AI, with only current Anthropic models scoring higher. Muse Wonder lands tenth on the same index.
Agentic work is where the gains are largest. On Terminal-Bench 4.0, the score moves from 33 percent to 57 percent, and on Terminal-Bench 4 Muse Wonder Full doubled its score while consuming 30 percent fewer output tokens. Coding also improves sharply: Muse Wonder Full reaches 68 points on the Coding Agents Index, tying Claude Code with Sonnet 5.5, while Muse Wonder scores 63, up from 54. Mathematics shows a similar pattern on ProofBench, a private benchmark for mathematical proofs, where Muse Wonder Full solves 94 of 100 problems, against 58 for its predecessor, leaving it five to six points behind the top models at 99 to 100.
Factual knowledge is the clear weak spot. On the AA-Omniscience knowledge test Muse Wonder Full scores 30 and Muse Wonder 25, well behind Claude Opus 5.5 at 46, GPT-6 Astra at 43 and Gemini 4 Argon at 42. The source reads this as a sign the Wonder series is built on a smaller base model than its competitors. Low hallucination rates partly compensate. Muse Wonder Full cuts its hallucination rate from 33 percent to 23 percent versus its predecessor, the second lowest among models scoring at least 45 index points; only Gemini 4 Argon, at 15 percent, hallucinates less often. Finance and legal performance stayed flat, dropping by up to two points.
Meta is pushing back on a familiar criticism that its Muse models are tuned to benchmarks rather than real-world use. In its blog post, the company says Muse Wonder Full performs better than expected on unfamiliar tasks that were not part of training optimization, and attributes that to a stronger base model. It also highlights two features aimed at agentic use. Active Memory lets the model retain key information across separate conversations and recall it when needed. Active Focus describes the model's ability to stay with difficult tasks, try alternative approaches, and ask follow-up questions when something is unclear. Meta frames both as central to a personal AI agent that can reach calendars, messages and documents.
On capability inputs, both models support a one-million-token context window and accept text, images, video and audio. Availability differs by variant: Muse Wonder is offered through Meta's API and Muse Code, while Muse Wonder Full is currently limited to Muse Code users. Muse Wonder will also power the Muse app soon, first for paying subscribers, and Meta says it plans to derive specialized models from it for all users over time. Meta states that the models were trained at its in-house Prometheus data center at gigawatt scale.
The economics are the sharpest trade-off. According to Artificial Analysis, Muse Wonder costs $4.52 per task on the Intelligence Index, nearly three times the $1.60 of Muse Spark 1.3. Token pricing rose to $4 per million input tokens and $20 per million output tokens, from $1.25 and $4.25 previously. Muse Wonder also consumes 28 percent more input tokens per task, and those input tokens make up 76 percent of total cost, which is why the per-task price rose so much faster than the headline rates. Both models do use about 10 percent fewer output tokens than their predecessor, and no token pricing is yet available for Muse Wonder Full.
In the wider market, that puts Muse Wonder below Claude Fable 5.1 at $7.63 per task and Claude Opus 5.5 at $5.98, but above GPT-6 Astra at $3.26, Gemini 4 Argon at $1.99 and GPT-6.1 Sol at $0.72. This is my own reading of the numbers: the strongest claim Meta can make here is capability at the frontier, not cost efficiency, and the pricing sits in the middle of a field that now spans a very wide range. A likely trade-off for teams is that cheaper models may still win on high-volume, latency-tolerant workloads where the extra agentic reliability of the Wonder series does not change the outcome. The flip side is that lower hallucination rates can matter more than raw knowledge scores in retrieval-augmented or tool-using pipelines, where the model is asked to synthesize supplied context rather than recall facts unaided. The one-million-token window and mixed-modality input also make these models plausible candidates for document-heavy agents, an area where input tokens dominate spend.
Why it matters: Meta is credibly back in the frontier conversation, which gives buyers a fourth serious family alongside Anthropic, OpenAI and Google, and the biggest gains land exactly where agentic coding tools compete. For developers and product teams, the practical question shifts from raw capability to cost per completed task, since a roughly 3x increase in task cost will reshape which workloads, and what level of autonomy, can be justified. Anyone currently using Muse Spark 1.3 should treat an upgrade as a capability-for-cost trade, and anyone reaching for Muse Wonder Full should note it is confined to Muse Code for now.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Meta's new Muse Wonder models rank in the top three on six out of ten benchmarks, with the biggest gains in agentic tasks. But the cost per task has nearly tripled. The article Meta's new frontier model Muse Wonder Full breaks into the AI top tier appeared first on The Decoder .