Muse Wonder vs. Claude Fable 5.1: Meta’s new model came closer than I expected
The New Stack Jessica Wachtel ● Covered by 25 sources
Meta’s Muse Wonder got close to Claude Fable 5.1 in tests. It was cheaper on paper, but token use and refusals changed the math.
Based on reporting by The New Stack, Jessica Wachtel — 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 launched Muse Wonder on October 12, roughly three months after Muse Spark 1.1 opened Meta’s paid Model API. The pitch is bold: Meta says the new series handles personal and work agent tasks well, leads its internal MemoryBench, and in the Full version tops coding benchmarks like DeepSWE and SWE-Marathon. That matters because Meta is putting Muse Wonder in the ring with Anthropic’s Claude Fable 5.1, one of Anthropic’s stronger models.
There are two Muse Wonder versions. The standard model is available through Meta’s Model API and Muse Code. Muse Wonder Full sits only inside Muse Code on the highest subscription plans. Pricing is where Meta starts to look aggressive: $4 per million input tokens and $20 per million output tokens, versus $10 and $50 for Fable 5.1. Cached input is $0.20 per million for Muse and $0.25 for Fable. On paper, that’s a real gap.
So I tested the standard Muse Wonder against Fable 5.1 on three tasks: a bug fix with hidden tests, an incident investigation, and a long-memory chat with a final deliverable plus quiz. Each model ran once at default reasoning. Meta’s own comparisons used maximum reasoning effort, which is not the same thing. The head-to-head results were mixed, but not in a trivial way.
On the bug fix, both models found the right files and fixed the harder issues without breaking regression tests. Fable passed all six hidden tests; Muse missed one because its fix still sent the old attempt count to a failure event. On the incident task, both scored 10 out of 10 and named the same config change as the root cause. Muse used more turns and more tokens in both cases, so despite the lower sticker price it ended up costing more in my runs.
The memory test is where Muse pulled away. It scored 25 of 25 on the deliverable and 8 of 8 on the quiz. Fable did not finish the task. Anthropic’s API returned a cyber refusal on turn 5, and Fable refused 34 of the 50 turns in that session, including the final deliverable and quiz. That leaves a weird split picture: Muse looked stronger at long-session memory, while Fable’s safeguards blocked what was basically a normal backend exercise. Across the full set, Muse came closer to Fable 5.1 than I expected, but it did not make the cheaper model the cheaper choice.
My take — AI-written commentary, not fact-checked reporting
This is the familiar AI pricing trick: slap a low token rate on the box, then let usage decide who really pays. Muse Wonder looks useful, but refusal behavior is the bigger story here than benchmark swagger. Teams do not buy model cards; they buy work getting finished, preferably without a safety system deciding a webhook relay is a crime scene.
Read more about this at: The New Stack