The leaderboard said frontier. The meter said stop.
14 Oct 2026 · News
Meta launched a near-frontier model. Within two days, the scores that mattered came from people's own work and their own quota.
Launch-day leaderboards no longer decide whether a model is good. Users decide within about 48 hours on two scoreboards the lab does not control: their own tasks and their own quota. At least one of the next two frontier launches will publish, on launch day, how much work a paid plan buys with the new model.
On 12 October Meta introduced Muse Wonder and Muse Wonder Full, models built for personal and professional agents, with Active Memory (carrying context across chats) and Active Focus (keeping at a hard goal through blockers). The launch page names "benchmaxxing" as the problem and leads with a playable 3D kitchen game. Wonder is on the Meta API and Muse Code; Wonder Full is for Muse Code subscribers.
Artificial Analysis put Wonder Full at 56 and Wonder at 52 on its Intelligence Index at max effort, behind Claude Opus 5.5 at 58. Wonder's cost per Index task landed at $4.52. The chart said near-frontier.
Within a day, people ran different tests. Paweł Huryn put Wonder Full against 105 real bugs frontier models had missed in early 2026 across two repos. Wonder Full fixed 44, a point behind GPT-6 Astra (45) and ahead of Fable 5.1 (43) and Opus 5.5 (41.7). Theo wrote that 3D game demos "don't represent real code work at all." Alexandr Wang, Meta's chief AI officer, agreed the team "did not explicitly train our model on 3d demo games" and asked people to "try muse wonder on your actual coding work."
The other scoreboard was the meter. Bijan Bowen posted that two Wonder Full results on the $50 plan used 108% of the five-hour limit and 36% of the weekly limit. Wang said the models were miscounting against Muse Code usage limits, and called it unintended behavior. By evening the bug was fixed and Meta reset everyone's usage. Users only knew the dial was wrong because it spun absurdly fast.
Twelve days ago we told readers that when a lab shows a leaderboard win, the habit to steal is asking what still fails on real work. This week the lab's own AI chief made that argument for his own model, after the crowd had already done it in public. Two days before Meta's launch we argued that an AI subscription is a meter you cannot see. Meta's meter was also broken. That is new evidence for the same point: a hidden dial can be wrong, and you find out only when it stops you.
Our call: of the next two frontier launches (the next Anthropic release, OpenAI's December answer, or Gemini 4 Argon going wider), at least one will publish on launch day how much work a paid plan buys with the new model: tasks, messages, or credits per window. The quota is now part of the product, because Meta just learned it is the first thing users test.
Test a new model on your own task. Watch what it does to your plan. Trust neither the chart nor the bill until both survive that check.