Meta’s frontier claim faces a subscription test

Meta has a new answer to criticism that its AI effort had fallen behind OpenAI, Anthropic, and Google: Muse Wonder and Muse Wonder Full. The two models, released on October 12, 2026, are Meta’s largest and most capable models to date, according to the company. They move away from Muse Spark’s low-latency focus toward expensive, long-running agent work.

Meta is putting the models behind API metering and paid tiers of Muse and Muse Code, asking users to pay $20 or $100 per month for the consumer agent and $50 or $200 per month for the coding product. The company has the distribution to put an assistant in front of billions of people. It has not established that an agent capable of navigating leases, insurance claims, calendars, messages, and code repositories is valuable enough for consumers and developers to subscribe.

Meta says Wonder is designed for personal and professional agentic tasks in law, finance, health, and software development. Its pitch is that the system can retain relevant details across sessions, recover after a failed approach, and request clarification before taking consequential actions. Those are practical failure modes once a chatbot is allowed to act rather than generate text.

Two models, three ways to pay

Muse Wonder is available now through the Meta Model API and Muse Code. Wonder Full, the larger variant, is restricted to Muse Code subscriptions; Meta is not serving it through the API. The company says Wonder will soon power more of its consumer Muse agent, with subscribers receiving access first, but it has not given a date for that rollout or described which capabilities will differ by subscription tier.

At $4 per million input tokens and $20 per million output tokens, Wonder is roughly equivalent in cost to GPT-6 Sol or Claude Opus 5.5, according to the reported comparison. That complicates Alexandr Wang’s claim that Meta’s developer offering will sit “meaningfully below” leading OpenAI and Anthropic models: the published API rate does not indicate a broad price undercut.

Recommended reading

Open weights give Mistral control, not frontier parity

Sergey Kuznetsov • Oct 6, 2026 • 9 min read

Wonder Full’s Code-only restriction draws a product boundary. Meta is making its most capable model a feature of a software-engineering agent subscription rather than offering developers raw API access. That may help the company package model inference with a higher-margin workflow product, but it gives API customers no route to the top model described in the launch materials.

The pricing arrives after the broader frontier-model market started cutting serving costs. In September 2026, we reported on lower-priced frontier-adjacent releases from Anthropic and OpenAI, including GPT-6 Sol and Claude Opus 5.5. Meta is entering that pricing contest with a model family whose core API rate is positioned alongside those products, not below them.

Active Memory and Active Focus are the product claim

Meta’s technical differentiation rests on two named capabilities: Active Memory and Active Focus. Active Memory is meant to carry information across conversations and retrieve the details relevant to the current task. Active Focus is meant to keep an agent pursuing a goal when it encounters a setback, try alternative approaches, recover from errors, and check with the user before following through on important actions.

Those descriptions are product behavior claims, not an architectural disclosure. Meta has not provided model size, training-compute figures, context-window length, inference latency, or the mechanism by which either capability is implemented in the supplied announcement. It is therefore not possible to determine whether Active Memory is persistent storage, a retrieval layer, a training-derived behavior, or some combination of those techniques. Nor has Meta said how users can inspect, edit, delete, or constrain the memories an agent carries between conversations.

The company’s framing aligns with its chosen evaluation measures. Meta highlighted MemoryBench, an academic benchmark intended to test whether language-model systems can continually learn from accumulated user feedback rather than merely answer questions over long documents. The MemoryBench paper argues that many existing memory evaluations overemphasize homogeneous reading-comprehension tasks with long inputs. Its authors simulate user feedback across multiple domains, languages, and task types, and conclude that the effectiveness and efficiency of current baseline systems remain far from satisfactory.

A strong MemoryBench score can support the claim that a system retains feedback better than another evaluated system. It does not demonstrate that a consumer agent safely handles a person’s mail, calendars, financial records, or private messages over months of use. Meta says its internal testing puts Wonder ahead of Fable, Opus, and Astra on MemoryBench, but neither the reported material nor the benchmark paper supplies independent evidence of reliability in those live-account scenarios.

Wang described the design target more plainly:

“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. 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.”

That is an appropriate objective for an agent. It is also where error recovery can become risky: an agent attempting alternative paths needs clear permission boundaries, not just persistence.

Benchmark position is strong, but not a complete comparison

On Artificial Analysis' Intelligence Index, Muse Wonder Full ranks behind Claude Opus 5.5 Max and ties Claude Sonnet 5.5 Max, according to the reported leaderboard results. Meta and third-party reporting also place the model ahead of or comparable with models including GPT-6 Astra and Claude Fable 5.1 in particular evaluations. Artificial Analysis said Wonder Full placed in the top three on six of the index’s 10 evaluations and was particularly strong in agentic knowledge work and scientific reasoning.

There are two caveats to the benchmark narrative. Meta’s release relies partly on internal testing, while the supplied reporting does not provide a full methodology or raw scores for all comparisons. Benchmark rank is also not the same as agent reliability. A model can perform well on coding or reasoning tasks yet still fail in the authorization, tool-use, and recovery loops required to operate across a user’s applications.

Google is the most obvious unresolved comparison. One report said Muse Wonder had leapfrogged Gemini 4, while another said the picture against Gemini 4 Argon remains unclear because that model’s capabilities are still coming into focus. The latter is the more defensible reading from the material available here: Meta has not supplied a settled, like-for-like comparison against Google’s newest system.

Meta has also made no claim in the supplied documents that Wonder Full is available through the API, and none of the supplied material provides independent latency, cost-per-completed-task, tool-success, or human-preference measurements. For developers choosing a production model, those omissions are at least as important as an aggregate intelligence ranking.

From Spark’s latency focus to a closed frontier tier

Meta says it began overhauling its AI work 14 months before the October 12 release. Muse Spark arrived in April 2026 as the first model from the reworked Superintelligence Labs, while Spark 1.3 arrived six weeks before Wonder. Spark was positioned as a lightweight, low-latency model; Wonder is explicitly larger and built for difficult tasks that can span a longer period.

This differs from Meta’s earlier identity as a supplier of open-weight Llama models. The company still says it intends to release larger open-weight models in the future and has released the smaller open-weight Muse Glimmer, but Wonder is a paid, controlled service. There is no disclosed download, self-hosting, or weight-release path for either Wonder model.

The strategic shift tracks Meta’s investment. The company invested $14 billion in Scale AI in June 2025, while another account put the finalized investment at $14.3 billion, and hired Scale cofounder Wang to lead its AI effort. Meta has said AI-related investment could reach $145 billion in 2026. The supplied reporting also says Wonder was trained at the Prometheus data center in central Ohio, which came online earlier in 2026, although Meta has not disclosed how much of the model’s training ran there or the compute used.

Meta reports more than 3.6 billion daily active users across its services, and Muse passed 5 million app downloads after launching last month. Those are distribution metrics, not evidence that users will accept a recurring paid tier. The pricing structure suggests Meta is trying to turn that reach into direct revenue rather than treating the agent as a free engagement feature around its advertising business.

Safety claims meet an authorization problem

Meta says it trained the Wonder family to avoid sharing private information without permission, refuse requests that facilitate harmful acts, and help improve cyber defenses without enabling abuse. It says the models are released under its Superintelligence Scaling Framework, but a full safety-evaluation report is still pending. The strongest safety assertions therefore lack the detailed evaluations needed to assess them.

“A personal agent can take on real tasks for you, which often means working across your messages, calendars, and documents. That’s why we trained the Muse Wonder family to avoid sharing private information without permission.”

The company is making those claims amid user complaints that Muse accessed personal messages without authorization. Meta disputes that characterization. Wang said some users may have authorized communications earlier in a conversation and later forgotten that authorization, while the company works to better explain product behavior.

That response points to a product-design problem rather than resolving it. If a user cannot readily tell which prior approval remains active, what data an agent can inspect, or which external action it can take, then a model’s refusal behavior is only one layer of protection. Meta says it is monitoring agent-related incidents or issues to understand and address problems, but the supplied material does not identify a permission model, an audit mechanism, or a user-facing control for Active Memory.

The distinction between defensive cybersecurity support and prohibited harmful assistance is important but underspecified. Meta says Wonder can help developers and security teams find flaws in their own code and strengthen systems while declining actually harmful requests. Without the promised safety report, there is no published account here of the evaluation thresholds, abuse testing, or cases in which the models are expected to refuse.

Meta has reached the benchmark conversation, not settled the agent case

Across the reports, Wonder is Meta’s strongest showing in the current model cycle. The pricing is consistent across the accounts: $4 input and $20 output per million API tokens for Wonder, $50 or $200 monthly for Wonder Full through Muse Code, and $20 or $100 monthly consumer Muse plans. The unresolved reporting issue is what the subscriptions deliver once Wonder enters the consumer agent, and when that will occur.

The sources differ in how definitively they frame Meta’s standing. One treats Wonder Full’s index placement as evidence that Meta has caught up with the frontier; another says the Gemini 4 Argon comparison is still unsettled. Meta has enough benchmark evidence to belong in the top-tier discussion, but it has not released the system-level evidence needed to show that its memory and focus claims make a personal agent dependable.

Meta can point to competitive scores, a large model API, and a large consumer funnel. But it is charging for an agent whose most valuable behavior—persistent work across private information and real-world services—also creates the hardest authorization and trust questions. The promised full safety report and missing detail on user controls over memory and permissions matter more than another leaderboard position.

Meta says Wonder will reach more of Muse soon, beginning with subscribers. It has not provided a concrete rollout date, an API route to Wonder Full, or the safety evaluation report it says will be available later.