Key Notes
- Meta’s Muse Wonder models combine Active Memory for relevant context across conversations with Active Focus for sustained work, error recovery and user check-ins.
- Meta’s model card reports strengths in legal work, memory and long-running coding, but mixes internal tests with third-party results and does not show a lead on every benchmark.
- Muse Wonder offers API access from $4 per million input tokens, while Wonder Full requires Muse Code’s $50 Power Usage or $200 Maximum Usage subscription.
Meta has launched Muse Wonder and Muse Wonder Full, two AI models aimed at agents that remember useful context and keep working through complex tasks. The company calls them its most capable models yet, positioning the October 12 release around coding, professional work and assistants that can act across applications.
The research announcement centers on Active Memory and Active Focus: carrying relevant information between conversations, then using it to pursue a goal despite errors or changing circumstances. The launch provides the product detail behind Zuckerberg’s claim that Meta has caught up with the leading AI labs, though its own benchmark table shows a more varied competitive picture.
Agents That Remember and Keep Working
Active Memory is designed to retrieve details from earlier conversations when they become useful again. For a personal assistant, that could mean carrying preferences or constraints into a later task instead of requiring the user to explain them repeatedly. The practical distinction is between storing past information and selecting the right information for the current job.
Active Focus addresses what happens after an agent starts working. Meta says the models can try different approaches, work around blockers and adapt to new information while seeking clarification or checking with the user before important actions. These capabilities target the extended workflows behind its Muse agent, where completing a task can require several tools and decisions.
In software development, that means tracking requirements through changes, tests and revisions rather than stopping after producing an initial answer. Meta’s developer materials also describe delegation to subagents and parallel tool use. The relevant test for users is whether that persistence produces a correct result, rather than simply a longer chain of activity.
What the Benchmark Results Actually Show
Meta’s model card reports strong results on legal-agent work, memory and sustained software engineering. Its evidence combines internal evaluations, third-party testing and results reported by competing model providers, so the figures should be read with their methodology rather than as one uniform comparison.
| Benchmark | Wonder Full |
|---|---|
| Harvey LAB-AA | 13.1 |
| MemoryBench | 85.5 |
| SWE-Marathon | 57.5 |
| AA Intelligence Index | 56 |
Higher is better within each row; the rows measure different things. MemoryBench and FocusBench are Meta’s internal evaluations, while the report attributes Harvey LAB-AA and the Intelligence Index to Artificial Analysis, and SWE-Marathon to Abundant AI. The SWE-Marathon comparison uses different agent systems for some competitors, and Wonder Full’s result uses xhigh reasoning effort.
The distinction changes the headline conclusion. Wonder Full scores 57.5 on SWE-Marathon, compared with 42.5 for GPT-6 Astra and 46.3 for Opus 5.5. Yet its Intelligence Index score of 56 falls below Opus 5.5’s 58, while the API-accessible Wonder scores 52. Wonder also slightly exceeds Full on MemoryBench, at 85.6 versus 85.5. Meta’s table supports specific strengths; it does not establish a universal lead over every competing model.
Wonder API Pricing and Wonder Full Access
Both models have a listed context window of 1M tokens. According to Meta’s developer page, Muse Wonder is available through the Meta Model API and Muse Code, with API pricing of $4 per million input tokens, $0.20 for cached input and $20 per million output tokens.
Muse Wonder Full is available through Muse Code’s Power Usage and Maximum Usage subscriptions, listed at $50 and $200 per month respectively. It is not offered as the same pay-per-token API option. That gives developers a different purchasing decision depending on whether they want to integrate Wonder into their own applications or use Full inside Meta’s coding product.
The release extends Muse Code beyond its earlier Spark-based launch. Meta also says Wonder will reach more experiences in the consumer Muse product, with subscribers receiving access first. That planned rollout is separate from the developer access available at launch.
More Useful Memory Also Requires More Trust
A persistent assistant can save users from repeating themselves, but its value depends on remembering accurate information and following current instructions. Old preferences can become outdated, and an agent that keeps working must still recognize when it needs permission or when the user’s goal has changed.
In its launch overview, Meta says it trained the models to protect private information and resist harmful requests while preserving defensive cybersecurity capabilities. For users, the next test is how reliably those safeguards and the new memory features work in everyday tasks, beyond the demonstrations and benchmark results accompanying the launch.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
