Meta’s Muse Wonder Matches Claude Fable for Less
Trending Topics Jakob Steinschaden ● Covered by 6 sources
Meta’s new Muse Wonder models just jumped up the A.I. charts. One is cheaper than Claude Fable and nearly matches it on quality.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta’s latest A.I. models, Muse Wonder and Muse Wonder Full, have made a sharp move up the rankings. Independent testing by Artificial Analysis calls it a “material jump” over the company’s earlier models, and the numbers back that up: Muse Wonder Full scores 55.8 on the Intelligence Index, while the standard Muse Wonder comes in at 52.4. That puts the full version just ahead of Anthropic’s Claude Fable 5.1, and the standard one only a point behind.
The speed of the improvement stands out almost as much as the ranking itself. Muse Spark 1.3, Meta’s predecessor, scored 48.1, and only six weeks separate the two releases. Meta is now sitting near the top of the chart with the kind of progress that usually takes much longer to show up, and it has done so without simply chasing raw price cuts.
Pricing tells the other half of the story. Muse Wonder costs $4 for a million input tokens and $20 for a million output tokens, up from $1.25 and $4.25 for Muse Spark 1.3. That makes it 2.8 times as expensive per test task as its predecessor. Still, it remains cheaper than Claude Fable 5.1, and the standard Muse Wonder costs about 40 percent less than Fable per test task. Muse Wonder Full is not in the price ranking at all because Meta does not charge per token for it.
The stronger model’s best work shows up in agent-style tasks, coding, and document-heavy tests. On Terminal-Bench 4.0 it jumps from 33 percent to 57 percent, matching Google’s Gemini 4 Argon. It also places first on CritPt and GDP.pdf, and scores 68 in Meta’s Muse Code environment, the same as Claude Code with Claude Sonnet 5.5. Hallucinations drop too, from 33 percent to 23 percent, though Gemini 4 Argon still does better at 15 percent.
But the shine is not uniform. Muse Wonder uses 7.69 million input tokens per task, far more than Google and OpenAI’s models, and about three-quarters of its cost comes from that input. On AA-Omniscience, its knowledge scores are weak next to the leaders. And access is split: Muse Wonder Full is only available to Muse Code users, while the standard version is the one exposed through Meta’s API.
My take — AI-written commentary, not fact-checked reporting
This is the part where the “cheaper” model is not actually cheap, just less expensive than the other expensive toys. Meta is selling progress, but also a bigger appetite and tighter access, which is classic platform behavior with an A.I. sticker on it. The real story is not that open models won; it’s that closed pricing still sets the mood for everyone else.
Read more about this at: Trending Topics