Tech • AI • Robotics • Game
Meta Just Dropped Muse Wonder (Better Than Opus 5.5?)
9.2/10TL;DR
Meta has launched Muse Wonder and Muse Wonder Full, agent-focused AI models aimed at personal and professional tasks, with benchmark gains in memory, legal work and long coding, but limited access and unresolved safety questions.
KEY POINTS
A new flagship for AI agents
Meta described Muse Wonder and Muse Wonder Full as its largest and most capable models yet, replacing Muse Spark as the company’s top-tier foundation for agent-style AI. The models were built specifically for personal and professional assistant work, including handling insurance claims, reading leases, and helping build business software. The launch marks the first major model release from a broader AI rebuild that began about 14 months ago.
Benchmark results put Meta closer to top rivals
In Meta’s comparison table, Muse Wonder Full was tested against Claude Fable 5.1, Claude Opus 5.5, and GPT-6 Astra. On a legal-agent benchmark, Wonder Full scored 13.1%, ahead of GPT-6 Astra at 8.6% and Opus 5.5 at 4.2%. On business workflow tasks, Wonder Full reached 45.4%, compared with about 41% for both Astra and Opus.
Strength in long coding, but not all coding
On long-duration coding projects, Muse Wonder Full posted 57.5%, outperforming Opus 5.5 at 46.3%. But the lead did not extend to every programming test: on building a full program from scratch, Opus 5.5 scored 70.5%, while Wonder Full scored 59%. Meta also showed its model trailing some rivals in computer-use evaluations and on the broader Artificial Analysis intelligence index, where Opus 5.5 held 58 versus Wonder Full’s 56.
Active memory is a central feature
One of the two headline additions is active memory, designed to retain important user context across conversations and retrieve it automatically when relevant. The goal is to let the agent remember preferences, budgets, and ongoing projects over days or months instead of forcing users to repeat details. On Meta’s own memory test, Muse Wonder scored 85.6%, while Opus 5.5 and GPT-6 Astra were both near 81%.
Active focus targets a common agent failure
The second major feature is active focus, which is meant to keep the model working through difficult tasks rather than stopping at the first error. Meta said the system can try alternate approaches, adapt to new information, request clarification when needed, and confirm with users before taking important actions. That directly addresses a common weakness in current agents, which often abandon tasks or falsely report them complete.
Demos point to cross-app automation
Meta showcased examples of the models planning a Saturday run in a browser across apps, designing and ordering custom gifts using social media photos, and creating a playable 3D kitchen game from images. These demonstrations reflect the broader push toward AI agents that operate across accounts, tools, and workflows rather than staying inside a single chat window. The examples remain vendor-produced demos, so independent real-world testing will matter.
Pricing is competitive, but access is split
For developers, Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, matching Opus 5.5 and undercutting Fable 5.1, listed at $10 and $50. It also offers a 1 million-token context window, with no added charge for long prompts. The catch is that Muse Wonder Full is not available through the API and can currently only be used in Muse Code on $50-a-month or $200-a-month plans.
Availability is limited for mainstream users
Muse Wonder is live in Meta’s API and in Muse Code, while Muse Wonder Full is restricted to paid Muse Code subscribers on higher-tier plans. The models are not yet broadly available inside the consumer Muse app. Meta said Wonder will be added to Muse experiences soon, with paying subscribers expected to receive access first, but no launch date was given.
Safety scores were better, not flawless
Because personal agents may access messages, calendars, and private documents, privacy handling is a central issue. Meta reported mishandling rates of 3.4% for Muse Wonder and 2.2% for Wonder Full, compared with 4.6% for GPT-6 Astra and 6.8% for Claude Fable 5.1. Those were the best scores in Meta’s chart, but 3.4% still implies errors in roughly 1 out of 30 tested cases.
CONCLUSION
Meta has moved closer to the front rank of AI models with an agent-first system that appears especially strong in memory, legal workflows, and long coding tasks. Whether Muse Wonder becomes a true rival to leading assistants will depend on independent testing, wider product rollout, and whether its safety performance holds up outside Meta’s own benchmarks.
Ask a question