[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space ● Covered by 10 sources
Meta shipped Muse Wonder, its new frontier model family. The bigger tier looks strong on agent tasks, but the launch also came with pricing, access, and quota drama.
Based on reporting by Latent Space — 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 Superintelligence Labs has finally shown the model family it has been building toward. Muse Wonder is the second-generation release, and Meta is pitching it as a frontier-class system for “personal and professional agentic work.” The family has two tiers: Muse Wonder, the faster version, and Muse Wonder Full, the more capable one. Meta says Wonder will power its personal-agent product, while Full is only available to Muse Code subscribers.
The launch had the usual pregame theater. In the hours before release, Meta insiders dropped watermelon hints, and leaked pricing circulated: $4 per million input tokens, $20 per million output tokens, and a 1M context window. Then the models went live, and users immediately hit a quota bug. Muse Code limits were draining too fast, Meta fixed it the same day, and quotas were reset.
On the substance, the early numbers are strong. Artificial Analysis put Wonder Full at 56 on its Intelligence Index v4.3, with Wonder at 52. That trails Opus 5.5 at 58, but it is a clear step up from Muse Spark 1.3’s 48. The biggest gains showed up on agent-style work. Terminal-Bench 4.0 jumped from 33% to 57% for Wonder Full, GDPval-AA v2.1 rose from 1681 to 1857 Elo, and AA-Briefcase moved from 1581 to 1729 Elo. Wonder Full landed in the top three on 6 of 10 evaluations.
Meta’s own headline claims are even louder. It says Wonder Full scores 57.5% on SWE-Marathon, 79.0% on DeepSWE v1.1, 13.1 on Harvey LAB, and 85.5 on MemoryBench. The company is also leaning hard on long-context and multimodal support: 1M context, text-image-video-audio input, and text output. But the pricing story is messy. Wonder’s published cost works out to $4.52 per index task, which sits below Fable 5.1 and Opus 5.5 but above Astra and the OpenAI models in the source table. Full has no published token price, so nobody can do the math cleanly yet.
There’s also a real split in reactions. Some users say Wonder Full is excellent for coding, agentic work, and long-horizon instruction following. Others think Meta is still a few months behind OpenAI and Anthropic, and complain that Full is locked behind a subscription with no public price. The launch reads like Meta’s classic move: show enough benchmarks to claim the frontier, then make everyone argue about whether the frontier is the right map.
My take — AI-written commentary, not fact-checked reporting
Meta is doing the modern AI equivalent of arriving late to a party and immediately turning up the music. The models sound legit, but the access model is already doing that annoying platform thing where the product is half demo, half gatekeeping. For a company this eager to sell “personal superintelligence,” the first test is simple: let people use it without a scavenger hunt.
Read more about this at: Latent Space