Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API

Meta is back in the race for the most powerful AI models, introducing two new proprietary models that the company says can outperform leading offerings from OpenAI and Anthropic on complex software engineering, legal work, business automation and long-running AI agent tasks.

Announced Monday, October 12, 2026, Muse Wonder and Muse Wonder Full are the latest releases from Meta Superintelligence Labs (MSL), the AI research organization led by former Scale AI CEO Alexandr Wang.

The launch marks a significant shift for Meta, whose earlier Llama models helped popularize downloadable AI models that organizations could run on their own infrastructure. Neither Wonder model is being released with publicly downloadable weights.

Instead, Meta is offering the standard Wonder model through a commercial API and reserving its most capable version for paid subscribers to its Muse Code programming assistant.

Through the API, the standard Wonder model costs $4/$20 per million input/output tokens, with cached input discounted to $0.20 per million tokens.

Independent benchmarking organization Artificial Analysis has already evaluated both models, awarding Wonder Full a score of 56 on its Intelligence Index — tying Anthropic's Claude Sonnet 5.5 and surpassing OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1, which each score 53. Only Anthropic's flagship Claude Opus 5.5 scores higher among these models, at 58. The standard Muse Wonder earns 52, placing it just one point behind GPT-6 Astra.

For enterprise buyers, the announcement introduces another serious competitor to established frontier AI providers — but also raises questions about cost, data governance and whether Meta's strongest technology will become available for production deployments outside its own tools.

From Muse Spark to a more ambitious class of agentic models

The release comes approximately 14 months after Zuckerberg reorganized Meta's AI efforts and established MSL, recruiting Wang and investing heavily in research talent and computing infrastructure.

In a post on X, Zuckerberg described Wonder as "a much larger and more advanced class of models" compared with the company's earlier Muse Spark family.

Meta says both Wonder models were developed using its Prometheus AI data center, combining large-scale initial training with reinforcement learning to improve their ability to complete complicated tasks.

The new models accept text, images, audio and video as inputs, produce text responses, and support context windows of approximately one million tokens.

Their primary selling point, however, is not conventional chatbot performance. Meta designed Wonder for tasks requiring agents to navigate multiple applications, retain information and recover from problems without abandoning the user's objective.

Meta identifies two capabilities as central to this approach:

1. Active Memory allows agents to preserve relevant personal and professional context across conversations, retrieve previous details and incorporate new information into ongoing work.

2. Active Focus is designed to keep agents pursuing objectives even when they encounter unexpected obstacles, changing requirements or failed attempts. Agents can revise their plans, explore alternatives and seek permission before consequential actions.

In its research announcement, Meta demonstrates agents developing software, organizing complex personal plans and working across web applications. The company also highlights professional uses including financial analysis, legal document preparation and insurance-related workflows.

Independent benchmarks place Wonder Full near the frontier

Artificial Analysis has published scores of 56 for Muse Wonder Full and 52 for Muse Wonder on its Intelligence Index, which combines ten evaluations covering reasoning, knowledge, coding and professional tasks.

The results put Wonder Full in a tie with Claude Sonnet 5.5 (56), ahead of GPT-6 Astra (53), Gemini 4 Argon (53) and Claude Fable 5.1 (53), but behind Claude Opus 5.5 (58). Wonder Full also substantially outperforms xAI's Grok 4.6, which scores 44 at high reasoning effort. All the scores are measured on the same index, with higher numbers representing better aggregate performance.

These are results from Artificial Analysis's model evaluations, not simply Meta's own composite scores. They also appear in Meta's launch comparison.

Meta has published additional results spanning agentic workflows, software engineering, professional knowledge work and memory.

Source: Meta's Muse Wonder Model Card, October 12, 2026. Scores are at maximum reasoning effort. Higher is better within each benchmark; scores across different benchmarks should not be compared directly. The table combines third-party results and Meta-run evaluations, as described in the model card.

The results suggest that Wonder Full is particularly competitive on sustained software engineering and certain professional agent tasks.

On SWE-Marathon, for example, Wonder Full scores 57.5 against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. On the legal benchmark, its 13.1 score substantially exceeds the compared models, although the relatively low absolute scores demonstrate how difficult it remains for agents to reliably produce fully satisfactory legal deliverables.

However, Meta does not lead across the board. Opus 5.5 remains stronger on Artificial Analysis's aggregate measure, while both Opus 5.5 and GPT-6 Astra outperform Wonder Full on Terminal-Bench 4.0. Opus also leads the comparison on health-related questions.

Meta's methodology deserves scrutiny. Its model card states that its comparison incorporates internally conducted evaluations, scores supplied by vendors and third-party results. Some software engineering comparisons use different coding tools, making them less controlled than testing every model under identical conditions.

The independent Artificial Analysis scores therefore provide an important counterweight to Meta's broader performance claims, but neither set of results replaces evaluations against an enterprise's own workflows.

API pricing starts at $4 per million input tokens

For enterprise developers, Meta is making Muse Wonder available through its Meta Model API, using the model identifier muse-wonder-1.0.

The model costs $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20 per million tokens.

Meta charges the same token rate regardless of how much of the available context window a request uses. Web search through the API carries an additional charge of $2.50 per 1,000 search queries.

The pricing places Wonder directly against Claude Opus 5.5 on per-token rates, while undercutting GPT-6 Astra and Claude Fable 5.1.

Sources: Meta's official pricing and Artificial Analysis pricing profiles for Opus 5.5, GPT-6 Astra and Fable 5.1. Prices are published list rates.

These figures matter for long-running agents that repeatedly process documents, invoke tools and generate extensive outputs. Such workflows can consume substantial token volumes even when their individual requests appear modest.

Lower token prices also do not necessarily translate into lower costs per completed task. Artificial Analysis estimates the cost of running its average Intelligence Index task at $4.52 for Wonder, compared with $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, reflecting differences in token consumption across models.

Meta's API documentation lists limits of 3,000 requests and four million tokens per minute, with a one-million-token context window. It also supports interfaces compatible with existing OpenAI and Anthropic software development tools, reducing the effort required to test Wonder within existing agent systems.

Compared to the other frontier or near-frontier models globally, Muse Wonder debuts in the upper-middle of the frontier pricing landscape: roughly twice the combined token cost of GPT-6 Sol, Claude Sonnet 5.5 and Gemini 4 Argon’s introductory rate, but still well below the $48–$60 premium tiers occupied by Opus 5.5 Fast, Claude Fable/Mythos and GPT-6 Astra.

That makes Meta’s positioning notable: Wonder is not competing as a bargain model like Muse Spark, but as a premium frontier offering priced alongside standard Claude Opus 5.5 and Gemini 4 Argon’s eventual full rate.

The chart is re-sorted from $0.30 to $120 total per 1M input + output tokens.

Wonder Full is restricted to paid Muse Code subscriptions

Meta is taking a different commercial approach with Wonder Full.

The company has not announced a pay-as-you-go API price for its most powerful model. Instead, Full is currently available through two Muse Code subscription tiers:

• Power Usage: $50 per month, providing access to Wonder Full and 20 times the usage allowance of the Everyday plan.

• Maximum Usage: $200 per month, providing 80 times the Everyday plan's usage allowance.

These subscriptions provide access through Muse Code, Meta's coding assistant, rather than unrestricted developer API usage. They also have usage limits.

That distinction is significant for enterprises seeking to embed a model in customer-facing software, internal automation platforms or production agent systems. Businesses can test Wonder Full within Meta's coding environment, but cannot yet deploy it through the standard Meta Model API.

The models are also proprietary rather than open-weight offerings. Unlike Meta's smaller Muse Glimmer model, which the company released under Apache 2.0, Wonder's parameters are not publicly downloadable for self-hosting, private infrastructure deployments or direct modification.

Meta's Model API Terms of Service permit commercial application development and give customers ownership of their submitted inputs and generated outputs, subject to the applicable terms. But access to Wonder itself remains governed by Meta's hosted service rather than an open-weight license.

Data training defaults and retention require careful review

An additional distinction could prove particularly important for enterprises handling proprietary source code, customer records or other confidential data.

Meta offers two data-use arrangements: Standard mode and Contributor mode.

Under Standard mode, Meta says submitted content is not used to train its models. The pay-as-you-go Wonder API is available exclusively in this mode.

However, Muse Code subscriptions default to Contributor mode, which permits Meta to use submitted prompts, code, files, contextual information and outputs to improve future models.

Subscribers can disable Contributor mode, but the change only applies to future submissions. Meta's terms explicitly prohibit submitting confidential, sensitive or personal information while using Contributor services.

Meta also offers zero data retention to qualified Model API organizations, subject to separate approval. This arrangement prevents ordinary storage of request and response content, with exceptions for legal compliance and misuse investigations.

There are operational trade-offs: Meta says zero-data-retention configurations do not support server-side conversation state, background requests, file uploads or built-in web search. Availability may also vary by model.

These limitations are particularly relevant to Wonder's advertised memory and long-running agent capabilities, which depend on maintaining useful information across extended workflows. Buyers will need to determine which features remain available under their required privacy configurations.

A new enterprise contender, with deployment questions remaining

Meta's technical model card also includes safety evaluations covering cybersecurity, biological and chemical risks, and the possibility of autonomous systems taking undesirable actions.

The company says it has applied additional safeguards and recommends human confirmation for consequential agent actions. Nevertheless, it acknowledges that the models can still make mistakes in unfamiliar multi-step tasks and that performance depends on the surrounding software and tools.

For enterprise technology leaders, Wonder represents more than another improvement in chatbot benchmark scores. Its combination of long-running agent capabilities, competitive third-party results and familiar usage-based API pricing offers another option for building systems that perform substantial work across applications.

But buyers face meaningful limitations: Wonder Full is unavailable through the public API, neither model can currently be self-hosted, and sensitive-data deployments require close attention to Meta's different service modes and retention rules.

The next test will be whether Meta can demonstrate that its benchmark improvements translate into more reliable, cost-effective automation inside real enterprise environments — and whether it makes its most capable model available beyond its own subscription-based tools.