Muse Wonder Full Leads Vals AI's Vibe Code Bench for Sustained Coding Work
Meta's top model scored 33.98% on Vibe Code Bench 1-100, which tests whether coding agents can extend a working app through up to ten dependent requests
By Icarus2 min read
Meta’s Muse Wonder Full leads Vals AI’s Vibe Code Bench 1-100, a coding benchmark that measures sustained, multi-turn development work rather than one-shot code generation. Wonder Full scored 33.98%, ahead of Anthropic’s Claude Opus 5.5 at 30.36% and Claude Opus 5 at 28.53%, according to Vals AI’s leaderboard. Meta AI chief Alexandr Wang highlighted the result on X on Wednesday afternoon.
The benchmark tests what most coding benchmarks skip: maintenance. It starts with a complete, working web application and hands the model an ordered sequence of up to ten product requests, written in plain English the way a product manager might send them. Each request adds, changes, or removes behavior, and later requests build on earlier ones. Every iteration has to add the new behavior while preserving everything that already worked; the first incomplete iteration ends the attempt.
Vals AI scores each model on the mean share of consecutive iterations passed across 50 test scenarios. The dataset holds 11,299 workflows, of which 3,747 (33.2%) are regression checks on earlier behavior. All models run on the same OpenHands harness with terminal access and preconfigured test services.
Behind Wonder Full, the leaderboard reads: Claude Opus 5.5 at 30.36%, Claude Opus 5 at 28.53%, Claude Fable 5.1 at 28.00%, OpenAI’s GPT-6 Astra at 27.64%, and the standard Muse Wonder at 26.6%. Meta’s previous generation, Muse Spark 1.3 Max, scored 20.5%.
Why It Matters
This is the second independent benchmark result favoring Meta’s new model family this week. Earlier Wednesday, DualEntry reported that Muse Wonder tied for second on its accounting benchmark. The Vibe Code Bench result points to a different claim: that Wonder Full can hold up over a long working session. That is the capability Meta emphasized at launch, with what it calls Active Memory and Active Focus.
Even the leader completes only about a third of consecutive iterations on average, and Vals AI notes that completing a full sequence of requests reliably remains difficult for every model tested. Benchmark scores also depend on harness and setup; Vals AI runs every model on the same OpenHands harness, which makes the comparison cleaner but does not guarantee real-world performance.
Developers weighing Muse Code’s paid tiers get the strongest outside evidence so far that the top tier holds up better than its rivals on long maintenance-style work. Wonder Full is available only through Muse Code’s Power Usage ($50 a month) and Maximum Usage ($200 a month) plans, not through the API. The field as a whole still fails most long arcs, but on this test, Meta’s model fails them least.
