The index score rose. Not every task did
The Decoder reported on October 12 that Meta released Muse Wonder and Muse Wonder Full. On Artificial Analysis Intelligence Index v4.3, Wonder Full scores 56 and Wonder scores 52. The previous Muse Spark 1.3 sat at 48.
On Terminal-Bench 4.0, performance went from 33 percent to 57 percent. Wonder Full places in the top three on six of ten benchmarks. That is a rank on those lists. It is not a claim that the model leads on every task.
[1]
Coding and proofs rose. Knowledge scores still lag
On the Coding Agents Index, Wonder Full scores 68. The Decoder writes that this ties Claude Code with Sonnet 5.5. Wonder scores 63, up from 54 for the previous model. On ProofBench, a private benchmark of mathematical proofs, Wonder Full solves 94 of 100 problems, up from 58. The report says the leading models are at 99 to 100. On Terminal-Bench 4, Wonder Full doubled its score while using 30 percent fewer output tokens. Finance and legal tasks stayed at the previous level, dropping by up to two points.
On the AA-Omniscience knowledge test, Wonder Full scores 30 and Wonder scores 25. The comparisons in the report are Claude Opus 5.5 at 46, GPT-6 Astra at 43, and Gemini 4 Argon at 42. Wonder Full’s hallucination rate fell from 33 percent to 23 percent against its predecessor. Gemini 4 Argon is at 15 percent.
On the Vals AI Index, Wonder Full (Max) debuts in fifth place, above the OpenAI models then on that list, with only current Anthropic models higher. Wonder is tenth. That is Vals’s ranking, not a rank Meta published.
[1]Where they are available, and why a task costs more
Both models support a one-million-token context window and take text, images, video, and audio. Wonder is available through Meta’s API and Muse Code. Wonder Full is currently limited to Muse Code users. Wonder will soon power the Muse app, starting with paying subscribers.
Meta highlights two abilities. Active Memory carries key information across conversations and recalls it when needed. Active Focus is sticking with a hard task, trying other approaches, and asking a follow-up when something is unclear.
Artificial Analysis prices Wonder at $4.52 per task on the Intelligence Index, against $1.60 for Muse Spark 1.3. Token prices are $4 per million input tokens and $20 per million output tokens, up from $1.25 and $4.25. No token price is available yet for Wonder Full. Both models use about 10 percent fewer output tokens than the previous model.
[1]