Meta Launches the Muse Wonder Series, Re-Enters the Frontier AI Model Competition

Meta has released two advanced models, Muse Wonder and Muse Wonder Full, designed primarily for long-duration agents, programming, computer operations, and specialized tasks in law, finance, and healthcare. They support up to 1 million tokens of context and accept text, images, video, and audio. The models introduce “active memory” and “active focus”: the former extracts task-relevant information from across conversations, while the latter enables the agent to try alternative approaches, recover from errors, and continue execution when encountering obstacles. Muse Wonder Full leads in the DeepSWE and SWE-Marathon long-range programming benchmarks but still lags behind some OpenAI or Anthropic models on Terminal-Bench, ProgramBench, HealthBench, and the overall intelligence index, and cannot be simply labeled as “comprehensively first.” The standard version is available via API at $4 per million input tokens and $20 per million output tokens; the more powerful Full version is accessible only to Muse Code subscribers. Meta has not disclosed model parameters, training data, or weights, and several results are based on internal benchmarks or different agent frameworks—further independent validation is still needed.

Article author, source: Meta

The Muse Wonder series includes two different versions. The standard Muse Wonder is designed for developers and can be accessed via the Meta Model API and Muse Code; Muse Wonder Full is a more powerful version, currently available only in the Power Usage and Maximum Usage subscriptions of Muse Code, with no standalone API pricing.

Both models support a 1 million token context, with input modalities including text, images, video, and audio, while output remains primarily text-based. Meta has not disclosed the model size, activated parameters, specific architecture, number of training tokens, or complete data composition, nor has it released downloadable weights. Therefore, this release represents Meta’s re-entry into the closed-source frontier model market, rather than an open-weight approach like Llama.

The model was trained at Meta’s gigawatt-scale Prometheus data center. The company states that over the past 14 months, it has rebuilt its AI technology stack from the ground up and expanded pre-training and reinforcement learning efforts based on Muse Spark. Reuters notes that this release also serves as a critical test of Meta’s significant investments in infrastructure, restructuring of its research teams, and competition for talent.

Meta refers to the two core capabilities of Muse Wonder as Active Memory and Active Focus.

Active Memory, or "active memory," emphasizes the model's ability to retain important context across multiple dialogue turns and proactively retrieve relevant information when needed for subsequent tasks. For example, when an agent plans an event, it can recall previously mentioned details such as budget, dietary restrictions, participant relationships, and scheduling without requiring the user to restate them.

This is not the same as a 1 million token context. The context window refers to the maximum amount of input that can be processed in a single inference; active memory, on the other hand, involves determining what information from longer-term accumulations is worth retaining and when to retrieve it. Meta has not disclosed its underlying storage format, retrieval algorithms, duration of memory retention, or how erroneous information is deleted; therefore, it is currently more accurate to view this as a combined capability of the model and agent system, rather than a fully disclosed new architecture.

Active Focus addresses the issue of agents giving up midway. When tool calls fail or webpage structures change, the model attempts alternative approaches, adjusts its plan based on new information, and asks the user for additional details when needed. For critical actions, it should first confirm with the user rather than acting directly to achieve its goal.

The official evaluation of such continuous execution capabilities uses FocusBench and SWE-Marathon. However, both FocusBench and MemoryBench are internal benchmarks from Meta, and neither the questions nor the full original outputs have been fully disclosed, making it currently impossible for independent researchers to fully reproduce them.

Muse Wonder Full excels in certain agent and long-range programming tests.

In DeepSWE 1.1, the Full version scored 79.0, while the Standard version scored 78.8; GPT-6 Astra scored 74.1 and Claude Opus 5.5 scored 74.2. In SWE-Marathon, which emphasizes longer software engineering tasks, the Full version achieved 57.5, outperforming the Standard version’s 53.1, GPT-6 Astra’s 42.5, and Claude Opus 5.5’s 46.3.

It scored 59.1 on Meta’s Personal Superintelligence Index, higher than the standard version’s 54.3, Astra’s 54.5, and Opus 5.5’s 55.2. On the AutomationBench-AA business automation test, the Full version scored 45.4, also outperforming the other compared models.

But it does not lead in all projects. In Terminal-Bench 4.0, the Full version scored 57.1, lower than Astra’s 59.1 and Opus 5.5’s 59.6; ProgramBench, which tests zero-shot programming, scored 59.0, significantly lower than Opus 5.5’s 70.5. On HealthBench Pro, the Full version’s score of 62.3 is also lower than Astra’s 65.3 and Opus 5.5’s 65.6.

The Artificial Analysis Composite Index gives the Full version a score of 56 and the Standard version a score of 52; during the same period, Claude Opus 5.5 scored 58. Therefore, a more accurate conclusion is: Muse Wonder Full has entered the frontier model group and excels in long-range programming, memory, and certain agent tasks, but has not yet surpassed all competitors comprehensively.

Model card disclosures include some results from third parties such as Artificial Analysis, Vals AI, DataCurve, and Abundant AI, while others come from internal Meta tests or self-reported results from competing companies.

The Agent frameworks used in different tests are not entirely identical. For example, GPT-6 Astra and Claude Fable 5.1 in SWE-Marathon run through Codex and Claude Code respectively, and not all models share the same tools, prompts, or execution environments. Meta also selects verified higher scores from internal results, official leaderboards, and vendor-reported outcomes.

The model card explicitly acknowledges that the prompts, tools, and runtime environments configured for competing models may not have been optimized to their fullest potential. This means the table can indicate general capability ranges, but small score differences should not be interpreted as stable or universal real-world advantages.

Meta also specifically addressed the质疑 about "benchmark gaming." The team compared the model's performance on post-training focus tasks versus tasks it was not specifically trained on, concluding that Muse Wonder Full's improvement on unfamiliar tasks exceeds the original trend of Muse Spark 1.3, attributing this to a stronger foundational model.

However, this generalization analysis was still designed and interpreted by Meta itself. The so-called “big model smell”—the phenomenon where models still appear intelligent on unfamiliar problems—currently lacks a unified, independently verifiable industry metric.

The API pricing for the standard Muse Wonder is: $4 per million input tokens, $0.20 for cached inputs, and $20 for outputs. The model identifier is muse-wonder-1.0, and the interface is compatible with the OpenAI SDK; existing applications can migrate by simply updating the API endpoint and model name.

Muse Wonder Full does not offer pay-per-token APIs and is available only through Muse Code subscriptions. The Power Usage plan costs $50 per month and provides 20 times the base tier usage; the Maximum Usage plan costs $200 per month and provides 80 times the base tier usage. Both plans have five-hour and weekly reset limits and are not equivalent to unlimited usage.

This product segmentation also complicates official comparisons: the most prominent achievements come from the Full version, but the version developers can directly integrate with is primarily the Standard version. Although the Standard version is close to the Full version, it falls short in the Personal Superintelligence Index, long-term execution, and certain coding tasks.

Meta plans to have Wonder power Muse, its personal agent, initially rolling it out to paying users and then developing specialized models for a broader audience. This indicates the company’s goal is not just to sell APIs, but to combine cutting-edge models, personal data integration, agent execution capabilities, and distribution channels like Facebook and Instagram.

Muse Wonder's individual agents may read emails, calendars, messages, and private files, and operate web browsers, desktop, and mobile applications. Active Memory, if required to retain context across sessions, also introduces issues related to data retention, false memories, and permission revocation.

Meta says this is the company's most aligned model to date, with enhanced protections against prompt injection, private information handling, cybersecurity threats, and chemical/biological risks.

The model card discloses that, without mitigation measures, Meta does not rule out Muse Wonder Full reaching a "high-risk" threshold in chemical and biological domains, thus incorporating safety post-training and system-level classifiers. Cybersecurity and失控 risks without mitigation were rated as "medium or lower"; after implementing safeguards, Meta rated the residual risks in all three domains as "medium or lower."

These conclusions still primarily stem from Meta’s own risk framework and internal governance procedures. The Agentic Misalignment test in the model card recording 0% harmful actions does not mean the model has no risk of going off-course in all real-world environments; Meta itself acknowledges that current tests cannot cover all scenarios.

The model may still make errors in novel multi-step reasoning, precise visual-spatial tasks, and languages with insufficient training coverage. Meta therefore recommends that, when an Agent is capable of performing irreversible real-world actions, deployers should additionally implement human confirmation and scenario-specific external safeguards.