October 13, 2026 | AI | By Mahesh

Meta Platforms (NASDAQ: META) released Muse Wonder and Muse Wonder Full on October 12, 2026, the first models its Superintelligence Lab positions as frontier class [1][6]. Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is limited to Muse Code subscribers. Neither model has downloadable weights, so Meta is competing with OpenAI and Anthropic on their terms: closed models sold by API and subscription.

What is Muse Wonder and what is new

Muse Wonder is Meta's new model family for personal and professional agent tasks, made up of Muse Wonder and the more capable Muse Wonder Full, both released October 12, 2026 [1][2]. It is larger, smarter and slower than Muse Spark, the lightweight agentic model Meta launched in April [7]. TechCrunch describes it as the first frontier model from Meta since Alexandr Wang restructured the company's AI effort in August 2025 [6].

Meta builds the pitch around two features. Active Memory retains what a user has shared across conversations, and Active Focus keeps a model working toward a goal, including checking in before important actions [1]. The launch demos include an insurance claim, a wedding seating chart and weekend run planning. Meta says Muse Wonder will power experiences in the Muse app, reaching subscribers first "soon" [1].

"a major leap toward personal superintelligence"

Meta, Official Announcement, October 12, 2026 [1]

Meta says both models match the best closed systems: its own comparison set is OpenAI's GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 [2][7]. Depth Grid covered the GPT-6 launch in Tech News Today, Oct 8.

Under the hood: context, modalities and training

Muse Wonder handles a context length of 1 million tokens at inference. It accepts text, image, video and audio input and produces text output only [2]. Meta has not disclosed a parameter count or an architecture, and the model card is silent on training data beyond saying it is multilingual with wide representation of languages [2].

Training took place at Prometheus, Meta's gigawatt-scale data center in central Ohio, using pretraining and reinforcement learning [1][6]. The API accepts requests in formats compatible with both OpenAI and Anthropic tooling, which lowers the cost of trying it in an existing agent stack [8]. The model id is muse-wonder-1.0.

Benchmarks: Meta's table against an independent index

Meta's model card reports results at maximum reasoning effort, with Muse Wonder Full's SWE-Marathon run at the higher xhigh setting. Meta also says competitor runs are best effort and may not be tuned for proprietary models [2]. The scores below come from that card as reproduced by VentureBeat [8]. Meta ran or selected them, and refusals count as zero.

The pattern is mixed. Muse Wonder Full leads on legal work, software engineering and long-running coding, for example 57.5 on SWE-Marathon against 42.5 for GPT-6 Astra and 46.3 for Opus 5.5. It trails both rivals on Terminal-Bench 4.0-AA and HealthBench Pro, and Opus 5.5 scores higher on the Artificial Analysis Intelligence Index, 58 to 56 [8]. Even the best legal score of 13.1 shows that reliable legal deliverables remain out of reach for agents.

Meta also says Muse Wonder beats Claude Fable, Claude Opus and GPT-6 Astra on MemoryBench, the test behind its Active Memory claim, but the launch coverage does not give the scores [6][9]. On cost, Artificial Analysis estimates an average Intelligence Index task at $4.52 for Muse Wonder, $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, so a lower token price does not guarantee a cheaper finished task [8].

Pricing, availability and how to access it

Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens through the Meta Model API, with cached input at $0.20 per million tokens and web search at $2.50 per 1,000 queries [4][8]. The rate is flat regardless of how much of the context window a request uses. Limits are 3,000 requests and 4 million tokens per minute. Opus 5.5 lists at the same $4 and $20, while GPT-6 Astra and Claude Fable 5.1 list at $10 and $50 [8].

Muse Wonder Full has no pay-as-you-go price and is not served through the API. It is included in Muse Code plans: Power Usage at $50 a month with 20 times the Everyday plan's allowance and Maximum Usage at $200 a month with 80 times [5][8]. Both carry usage limits.

Buyers should read the data terms. Muse Code subscriptions default to a Contributor mode that lets Meta use submitted content to improve its models, and confidential or sensitive material is prohibited in that mode. Zero data retention configurations do not support server-side conversation state, background requests, file uploads or built-in web search, which are the features that long-running agents and memory depend on [8].

Limitations, trade-offs and alternatives

Neither model can be self-hosted, and the best one cannot be bought by API. The model card lists known weaknesses: multi-step reasoning errors in novel situations, weak visual and spatial reasoning (3D tasks in particular) and possible drops in less-supported languages. It says the models are not intended for people under 18 and recommends human confirmation before irreversible agent actions [2].

Meta's own risk assessment is candid. Without mitigations it does not rule out reaching its "high" threshold in chemical and biological risk, so it added safety post-training and system-level classifiers, after which all domains rated moderate or lower [2][3]. Muse Wonder Full scored 88.4% on the CyberGym test. Meta published a framework for these thresholds and promises a full safety report soon [1]. Washington has also ordered AI firms to report incidents, a rule Depth Grid covered in White House orders AI firms to report incidents after Anthropic.

The business case is still an open question. Fortune notes Meta's capital spending of $145 billion against annual revenue of $201 billion, of which digital advertising is about 98%. Deutsche Bank analyst Benjamin Black speculates that a personal AI agent could reach up to 8% of revenue by 2030, about $36 billion, and Fortune doubts that would cover the cost of AI engineers [7].

Teams that need weights have another route. Mistral's open-weight model is due by the end of October, and Depth Grid has its specs in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing. For closed models, Opus 5.5 leads on the independent index and GPT-6 Astra has the lowest estimated cost per task.

Sources

1. Meta. "Muse Wonder and Muse Wonder Full," October 12, 2026. research.meta.ai
2. Meta. "Muse Wonder Model Card," October 12, 2026. research.meta.ai
3. Meta. "Meta Superintelligence Scaling Framework." research.meta.ai
4. Meta. Meta Model API developer portal. dev.meta.ai
5. Meta. Muse Code plans and access. dev.meta.ai/code
6. TechCrunch. "Meta releases its first frontier model, Muse Wonder," October 12, 2026. techcrunch.com
7. Fortune. "What Muse Wonder and frontier AI mean for Meta," October 13, 2026. fortune.com
8. VentureBeat. "Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API." venturebeat.com
9. MemoryBench paper, arXiv. arxiv.org

Read More on Depth Grid

• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Tech News Today, Oct 12: Anthropic Report, Starlink, Apple
• Tech News Today, Oct 8: GPT-6, SpaceX, Nuclear Clocks
• Nvidia in Talks to Deepen Stake in or Buy Reflection AI

Article by Mahesh | Depth Grid