TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Meta Launches New Model Muse Wonder, Zuckerberg Joins the Frontier AI Competition

Meta Launches New Model Muse Wonder, Zuckerberg Joins the Frontier AI Competition

2026.10.13
Share
On Meta's Personal Superintelligence Index, Wonder Full scored 59.1, and Claude Opus 5.5 scored 55.2.

Author: Decrypt

Translated by: TechFlow

TechFlow Brief: This time, Meta isn't using open source as a pretext; instead, it goes head-to-head with OpenAI and Anthropic's most powerful models, and proactively discloses Claude's poor performance in select tests. For industry professionals focused on AI costs, model capabilities, and data privacy, this launch serves as both a signal of a price war and a stress test regarding the boundaries of “memory-equipped assistants.”

On Monday, Meta launched Muse Wonder and Muse Wonder Full, the first frontier-grade models from Meta’s Superintelligence Lab, boasting a 1 million-token context window.

In Meta’s own charts, Wonder Full outperforms GPT-6 Astra and Claude Opus 5.5 in long-range programming, legal agent tasks, and memory. However, independent rankings from Artificial Analysis still place Opus 5.5 first with a score of 58, while Wonder Full scores 56.

The API version of Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens, matching Opus 5.5, and is lower than the $10 and $50 rates for GPT-6 Astra.

On Monday, Meta released Muse Wonder and Muse Wonder Full, the inaugural models from its Superintelligence Lab, signaling the company’s willingness to pit them head-to-head against OpenAI and Anthropic’s top-tier offerings.

"Muse Wonder represents a major step toward providing personalized superintelligence for everyone," Zuckerberg wrote on Threads. "Congratulations to everyone on our team for reaching this milestone. I'm excited to see how you put it to use."

This phrasing safely aligns him with President Trump. On Thursday last week, Trump stated on Truth Social that the White House considers anyone who says "artificial intelligence" rather than his preferred term "superintelligence" to be an "enemy."

Since 2025, Meta has referred to its AI division as Meta Superintelligence Lab, meaning Zuckerberg was already compliant before it became a “mandatory requirement.”

Meta’s launch highlights two key features: proactive memory, which carries information shared with the model across successive conversations; and proactive focus, enabling the model to persistently tackle difficult objectives, experiment with new approaches, and consult you before taking any major actions.

In April 2025, Meta’s open-source Llama 4 model arrived with high promises but quickly faltered in independent testing, as doubts emerged that versions scoring well on popular public leaderboards behaved differently from those available for download.

Zuckerberg subsequently acquired Scale AI in a deal worth approximately $14 billion, placing its founder Alexandr Wang in charge of a new division named Meta Superintelligence Lab. Its inaugural model, Muse Spark, was released in April as a small, fast, closed-source model, also marking the end of the open-source Llama lineage.

Muse Wonder’s release comes just weeks after OpenAI launched GPT-6 Astra on September 3. Anthropic followed suit on September 22 with Claude Opus 5.5, claiming it can match the more expensive Claude Fable 5.1 across most workloads.

On Meta’s Personal Superintelligence Index, Wonder Full scores 59.1. This index provides an internal composite score for personal and professional assistant tasks, where higher is better, though the exact methodology remains undisclosed. Claude Opus 5.5 scores 55.2, GPT-6 Astra scores 54.5, and Claude Fable 5.1 scores 49.9.

In the SWE-Marathon benchmark measuring AI coding capabilities, Meta’s models performed well. Wonder Full scored 57.5, compared to 46.3 for Opus 5.5 and 42.5 for Astra. However, Meta acknowledges that this comparison was not a controlled experiment, and its models were run in maximum effort mode.

The Harvey legal agent benchmark assigns over 1,200 lawyer-style tasks to AI models, counting only fully completed works that pass every check on an expert scoring rubric, with no partial credit awarded. Wonder Full scored 13.1, Astra scored 8.6, and Opus 5.5 managed only 4.2.

On MemoryBench, both Wonder models scored around 85.5. This is Meta’s proprietary test measuring a model’s ability to retain facts within conversations spanning 64,000 to 100,000 tokens, including details later updated. Opus 5.5 scored 81.5, while Astra scored 81.4.

These figures come from Meta, and two of the tests were developed internally. Independent rankings from Artificial Analysis, which compute a weighted average of 10 evaluations covering AI agents, programming, general knowledge, and scientific reasoning, still rank Opus 5.5 first with a score of 58, tying Wonder Full with Anthropic’s cheaper Claude Sonnet 5.5 at 56.

The commercially available Muse Wonder for developers scores 52 on this index, one point below Astra and Fable 5.1. Priced at $4 per million input tokens and $20 per million output tokens—with tokens referring to word fragments read and written by AI models—this pricing exactly matches what Anthropic charges for Opus 5.5 and is significantly lower than the $10 and $50 OpenAI charges for Astra.

For everyday users, the “memory” feature is a double-edged sword, as an assistant that remembers everything holds vast amounts of personal information, and Meta’s track record with private data leaves much to be desired. Last month, Inc. columnist Jason Aten discovered that Meta’s Muse agent continued syncing over 187,000 lines of his private iMessage history even after he denied access, subsequently claiming it had only seen notification previews.

Meta now refers to Wonder as “our most aligned model to date,” with model cards showing a 0% failure rate in agent misalignment tests. These tests stress the model and tally the frequency of harmful actions taken. This comes barely two months after Meta disclosed that a Muse Spark model breached its sandbox during safety evaluations and compromised a third-party service.