Muse Wonder AI models add memory to follow your preferences

The Muse Wonder AI models add two features aimed at making AI agents, the kind that carry out multistep tasks for you, more dependable.

Active Memory lets the model learn your preferences and carry relevant context from one chat into the next. Today, many chatbots start each conversation with a blank slate. With Active Memory, if you told it last week how you like emails written or which bank you use, it can pick that up again.

Active Focus is a persistence loop. When the model’s first approach fails, it tries another, recovers from errors and keeps going until the task is finished. Meta says it also checks in with you before taking important actions. The aim is to keep the model locked on a goal during long jobs, such as a big research project or a large piece of software, where AI often drifts or piles up compounding mistakes.

The results are strong, with caveats. On the Artificial Analysis Intelligence Index, an independent score that combines many tests, Muse Wonder Full scored 56. That ties Claude Sonnet 5.5 and beats GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, which all scored 53. Only Claude Opus 5.5 ranked higher, at 58. The standard Muse Wonder model scored 52.

Wonder Full shines on long coding jobs. On a test called SWE-Marathon it scored 57.5, against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. Meta also says it leads on some specialist tests, including work with legal documents.

It does not win everywhere. Opus 5.5 and GPT-6 Astra both beat it on Terminal-Bench 4.0, a test of working in a computer’s command line, and Opus leads on health questions. When the code editor Cursor added Muse Wonder a few days after launch, it scored 46.1% on CursorBench in its High setting, or up to 50.4% at Max, solidly mid-pack next to Claude Opus 5.5’s leading 57.8%. Benchmarks are also not a perfect guide to how a model feels in daily use.

This is an incredible form by @AIatMeta pic.twitter.com/fTcUKkvnqu

— Digital Trends (@DigitalTrends) October 12, 2026

Meta says it taught the new models not to reveal people’s private details unless they agree, to refuse requests that could cause real harm and to help strengthen defenses against cyberattacks. It says Wonder has its strongest protections yet against chemical, biological, nuclear and cyber misuse, under a new Meta Superintelligence Scaling Framework.

That focus comes after a string of incidents in which AI agents from rival labs behaved in unexpected and risky ways when given open-ended instructions. Mark Zuckerberg said in September that Meta had delayed the Muse models to work on safety. Meta has not said whether Muse Wonder agents misbehaved in its own testing.

For now, the most capable model is mostly in developers’ hands. Muse Wonder is available through the Meta Model API and Muse Code, Meta’s AI tool for writing software, and reportedly on OpenRouter. Developers pay $20 for each million tokens (chunks of text) the model produces, and $4 per million it reads in, the same rates as Claude Opus 5.5.

The bigger Wonder Full model is not offered through any public API. Access is limited to paying Muse Code subscribers on the Power and Maximum plans, and reports differ on what those plans cost, so check Meta’s own pricing page before signing up.

Everyday users will see it more gradually. Meta says Wonder will soon power more of its Muse app, starting with subscribers, and will be used to build specialized models for all Muse users later. The models can handle a huge amount of material at once (a context window of 1 million tokens) and accept text, images, video, audio and PDFs.

If memory across chats sounds useful but a little unnerving, it is worth checking what your AI app stores and how to clear it once these features reach you.

Meta was once a leader in AI with its Llama models, which anyone could download and run. It then fell behind after several disappointing releases while Anthropic, OpenAI and Google raced ahead.

Zuckerberg responded last year by pouring billions into a new group, Meta Superintelligence Labs, led by former Scale AI chief executive Alexandr Wang. Meta says the rebuild began 14 months before this launch, and that both new models were trained at its gigawatt-scale Prometheus data center. Its first model from the new lab, Muse Spark, arrived in April, followed by the Muse personal agent in September.