Meta Launches Muse Wonder Models as Paid AI Push Meets a $200 Billion Spending Forecast

Two new models put Meta's AI spending to the test

Meta launched Muse Wonder and Muse Wonder Full on October 12, 2026, claiming a return to the leading ranks of artificial intelligence while putting its strongest new capabilities behind paid access. The release connects three parts of its strategy: more capable personal assistants, revenue from subscriptions and developers, and a computing expansion that could become substantially more expensive.

The company expects 2026 capital expenditure, including principal payments on finance leases, of $130 billion to $145 billion. Financial services firm Cantor Fitzgerald forecasts that spending could reach $200 billion in 2027. That larger figure is an outside forecast, not an announced Meta budget. It would represent an increase of about 38% from the top of this year's guidance, or about 54% from the bottom.

The models arrive after a disappointing Llama 4 release in 2025 and roughly 14 months of rebuilding Meta's AI operation. They target advanced programming and tasks involving legal documents, financial information and health information. Meta also wants them to power agents, software that can use tools and take actions rather than simply answer questions.

There is evidence of stronger performance, but not a clean sweep. Meta's published software development results put both models above Anthropic's Claude Opus 5.5 on one benchmark. A separate Artificial Analysis index places Muse Wonder Full below that rival. The immediate business test is whether those capabilities persuade people to keep using, and paying for, Meta's products.

What is available now, and what comes later?

Muse Wonder is available to developers through the Meta Model API and the Muse Code programming assistant. Muse Wonder Full, the more capable version, is restricted to Muse Code subscriptions. Neither model is available for public download or for users to run on their own servers.

The personal assistant Muse is a separate product. Its app is free, with paid plans offering additional usage. Meta plans to introduce Wonder into that assistant first for subscribers, but has not announced a firm rollout date or a schedule for wider access.

The sequence matters because a model launch is not the same as an upgrade reaching every consumer. The main developments are:

• 2025: Llama 4 fell short of expectations, prompting changes to Meta's AI organization and hiring.
• September 2026: Meta introduced Muse Spark and the Muse personal agent. Spark was the first model from the rebuilt Meta Superintelligence Labs division.
• Late September 2026: Amazon blocked Muse from making purchases on behalf of users.
• October 12, 2026: Muse Wonder and Muse Wonder Full became available through the announced developer products.
• Next steps: Wonder is due to reach Muse subscribers, while more specialized models and a fuller safety report are planned without published dates.

The new models build on Muse Spark. Meta's expanding lineup also includes Muse Image, although the Wonder launch centers on agents and programming rather than image generation.

Memory and persistence are the main product claims

Meta describes two central capabilities: Active Memory and Active Focus. Active Memory retains relevant information from earlier interactions and retrieves it for later tasks. Chief AI Officer Alexandr Wang says that information can span days, weeks or months, allowing an assistant to carry preferences and requirements across conversations.

Active Focus addresses what happens when a task stops going to plan. Meta says the models can recover from errors, change their approach after a tool fails, seek clarification and request confirmation before important actions. These are intended to reduce situations in which an agent gives up or repeatedly follows an unsuccessful method.

Wang, who leads Meta's AI operation, explained the design goal in terms of the obstacles that arise during everyday tasks:

When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks," Wang said. "We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time.

Meta points to insurance claims, lease interpretation, personal finance and advanced programming as potential applications. Muse already targets activities such as grocery shopping, travel bookings and dinner reservations. These examples describe intended uses, not proof that the models complete each task accurately or safely.

Both models reportedly accept images, video and audio and have a context window of one million tokens. Tokens are the units into which a model divides information for processing. A context window measures how much it can consider at once; persistent memory is a separate capability for carrying selected information between interactions.

The benchmark picture is competitive, not conclusive

On SWE-Marathon, a benchmark involving substantial software development tasks, Meta's published results give Muse Wonder Full a score of 79%, Muse Wonder 78.8% and Claude Opus 5.5 74.2%. Those figures put Full 4.8 percentage points ahead of the Anthropic model and standard Wonder 4.6 points ahead.

The gap between Meta's own models on that test is only 0.2 percentage points. That suggests buyers should examine the tasks they actually need performed rather than assume the larger version delivers a large advantage in every setting. The results are Meta's published comparisons, not a guarantee of performance on an organization's software projects.

The Artificial Analysis Intelligence Index gives a different view. Muse Wonder Full scores 56, while standard Wonder scores 52. Claude Opus 5.5 leads those models with 58, and OpenAI's GPT-6 Astra scores 53. Accounts describing the rankings place Full in a tie for second and Wonder in a tie for seventh; one identifies Claude Sonnet 5.5 as sharing Full's score.

One account calls the larger Meta model Muse Wonderful, but the other launch accounts identify it as Muse Wonder Full and give the same score of 56. The naming discrepancy should not be mistaken for evidence of a third released model.

Meta also cites comparisons with Claude Fable 5.1 and GPT-6 Astra. Its position against Google's Gemini 4 Argon is less settled: one reported index gives Gemini 53, while another account says the broader comparison is still unclear. Different tests measure different abilities, so a strong programming result cannot establish superiority across legal analysis, financial tasks or personal assistance.

Subscriptions and API charges are different costs

Muse's consumer subscription plans cost $20 or $100 per month. Muse Code plans cost $50 or $200 per month and provide access to Wonder Full. Those subscription prices should be kept separate from Wonder's API rates, which charge developers according to the amount of information processed.

The reported API price is $4 per million input tokens and $20 per million output tokens. Input is the information sent to the model; output is what it generates. At those rates, a workload using one million input tokens and one million output tokens would cost $24, before any applicable discounts or other charges. That example is a calculation from the published rates, not a subscription price.

Cached input can receive discounted pricing, although no specific discounted rate is given. Caching allows previously processed material to be reused rather than handled as entirely new input. For an agent that repeatedly consults lengthy documents, that distinction can affect costs.

Wang says Meta's API will be priced meaningfully below leading OpenAI and Anthropic models. The reported price comparison is narrower: Wonder's rates are described as matching Claude Opus 5.5 and amounting to about 40% of GPT-6 Astra's pricing. That supports a claimed discount against the OpenAI model, but not against the named Anthropic model at the quoted rates.

Reports also differ on availability. Some describe API access as planned, while the more specific launch accounts say standard Wonder is already available through the API. No API access has been announced for Wonder Full. Although the strongest version has been described as aimed at direct enterprise sales, its concrete launch channel is Muse Code.

A more capable model still needs access to websites

Muse reportedly reached nearly nine million global downloads after its September launch, according to Sensor Tower data. Downloads measure initial interest, however, not lasting engagement. Its download growth has slowed, and its retention after 14 days is reported to be below ChatGPT's. Exact retention percentages have not been given.

Some of the obstacles are outside Meta's model. Jefferies analyst Brent Thill said websites blocking agents were damaging the user experience. Amazon's decision to block Muse purchases at the end of September illustrates the problem: an assistant can understand a shopping request and still be unable to complete it.

Active Focus may help when a tool fails or information is missing. It does not establish that an agent has permission to bypass a website's restrictions. Successful personal assistance depends on cooperation from external services as well as reasoning, memory and persistence.

Price also complicates the case for using the most powerful models everywhere. OpenRouter platform figures show the top ten models described as high quality accounting for 14% of usage in the third quarter of 2026, down from 55% in the first quarter of 2025. That is a decline of 41 percentage points on that platform, not a measurement of the entire AI market.

The pattern suggests that cheaper models are taking a larger role in routine work. Meta is developing specialized versions intended to preserve much of Wonder's capability while operating more efficiently, which could help it serve a larger consumer audience without using its largest model for every request.

Privacy depends on both permissions and product settings

Agents operating across messages, calendars and documents need access to sensitive information. Meta says Wonder is trained to avoid sharing private data without permission and to refuse assistance with harmful conduct. Reports also describe the models as seeking confirmation before consequential actions.

Some users have complained on social media that Muse accessed personal messages without authorization. Meta disputes those complaints. Wang says the company is monitoring agent incidents and explaining permissions to users, including cases in which someone may have authorized communications earlier in a conversation and later forgotten.

That disagreement leaves the specific complaints unresolved in the public accounts. Persistent memory and the ability to act across services make clear permission controls especially relevant: remembering information is different from receiving permission to use or share it.

Data handling also varies by access route. Muse Code subscriptions reportedly default to Contributor mode, which permits submitted material to help improve future models. Subscribers can disable that setting. The standard Model API reportedly does not use submissions for training. These are reported product conditions, and they should not be treated as identical privacy arrangements.

For businesses, this distinction is separate from benchmark quality. A coding assistant may perform well while still requiring a review of whether proprietary code, client documents or personal information can be submitted under its selected settings.

Safety claims await a fuller evaluation

Meta says Wonder provides its strongest safeguards so far against cyber misuse. Those protections are currently associated with the paid model, rather than described as available across every free Muse experience. The statement does not establish that unpaid products have no safeguards; it distinguishes the new model's claimed protections.

The company says development follows its Meta Superintelligence Scaling Framework and that a fuller safety evaluation report is forthcoming. It has not given a publication date. Without that report, the detailed assessment behind the new safety claims remains incomplete in public.

Chief Executive Mark Zuckerberg argues that AI can be controlled through safeguards and responsible development. He has rejected calls for the entire industry to slow research, while acknowledging that Meta has slowed some internal development to address product safety.

Zuckerberg described his position on control and autonomy:

"My view is much closer to the belief that AI is a technology that can be controlled through safeguards and built responsibly, rather than the view that it will develop a sense of self and act on its own," he said.

He did not give an estimate of the probability of catastrophic harm. His position places responsibility on individual laboratories to manage safety while continuing development. The forthcoming evaluation is therefore a consequential missing piece, particularly for models designed to act rather than merely respond.

The computing expansion reaches beyond Meta

The Wonder models were trained using Meta's Prometheus data center. The launch follows investment in infrastructure, recruitment and an organizational rebuild under Wang. Meta invested about $14 billion in Scale AI and brought Wang into its leadership as part of that effort.

One account describes Wonder, internally codenamed Watermelon, as using roughly an order of magnitude more computing resources than Muse Spark. That is a reported comparison, not a published accounting of training costs or hardware requirements. It is consistent with Meta's argument that reaching the leading ranks required substantially more computing capacity.

Zuckerberg says the next stage of model development is already underway and that the pace will accelerate. Meta is working on larger models as well as more efficient specialized versions, pursuing greater capability and lower operating costs at the same time.

The spending outlook creates potential demand for server manufacturers associated with Meta's AI supply chain, including Hon Hai, also known as Foxconn, Quanta, Wistron, Wiwynn and Inventec. The expectation that they could benefit rests on continued infrastructure purchases. No new contract values, supplier allocations or delivery schedules have been announced in connection with Wonder.

Capital expenditure is also broader than a single model's training budget. Meta's $130 billion to $145 billion guidance includes finance lease principal payments, and Cantor Fitzgerald's $200 billion forecast concerns the following year. Neither figure should be read as the price of developing Muse Wonder alone.

Investors are responding before revenue is proven

Meta shares closed the launch session at $742.96, up 3.38%. Other accounts round the gain to 3.4% or approximately 3%. The stock had already begun rising before Zuckerberg's afternoon announcement, so the full daily move cannot be assigned solely to the model release.

The broader stock figures cover different periods: one account gives a 13% gain for the year and 31% since early September, while another reports nearly 10% since Muse's launch later that month. Those numbers are not directly comparable without matching their starting dates.

Bank of America securities analyst Justin Post has argued that investment in Meta Superintelligence Labs is translating into better model performance and product innovation. The new release supplies more evidence for that view, but does not yet establish how much revenue the models will generate.

The commercial questions are specific. Meta has not disclosed how many Muse users will pay for Wonder, how many developers will choose its API, or how the revenue compares with the cost of serving persistent agents. Broader consumer availability, API access for Full and the safety report also lack announced dates. The launch strengthens Meta's competitive position while leaving those business and deployment decisions open.

Key Points

• Meta launched Muse Wonder and Muse Wonder Full on October 12, 2026, with developer access first and a Muse consumer rollout still to come.
• Meta's programming benchmark favors its models, while Artificial Analysis places Wonder Full below Claude Opus 5.5.
• Wonder's API costs $4 per million input tokens and $20 per million output tokens. Full requires a Muse Code subscription.
• Neither model is available for public download or operation on users' own servers.
• Memory, permissions, website restrictions and reported training settings remain practical issues for agent deployment.
• Meta's 2026 spending guidance is $130 billion to $145 billion. The $200 billion figure for 2027 is Cantor Fitzgerald's forecast.
• A fuller safety evaluation and broader availability dates have not been published.