Meta introduced Muse Wonder and Muse Wonder Full on 12 October 2026, calling them its most capable models to date and saying they deliver state-of-the-art performance on personal and professional agentic tasks. Mark Zuckerberg announced the family on X the same day.
Meta says Muse Wonder is available today through the Meta Model API and Muse Code, that Muse Wonder Full is available today for Muse Code subscribers, and that the models build on the groundwork it laid with Muse Spark.
Business Pill · BENCHMARKS
A one-minute explainer of benchmarks: why a test score is not the same as your own result. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, Meta is pitching Muse Wonder on agent work rather than chat: its own table puts the models on top for its personal-agent index and long-horizon coding, while Claude Opus 5.5 scores higher on several coding and general rows.
Meta says it began a ground-up overhaul of its AI efforts fourteen months ago, and that the new models build on that work across pretraining and reinforcement learning. According to its research blog, both were trained at Meta’s gigawatt-scale Prometheus data center.
The model card lists a context length of 1M tokens at inference, with text, image, video and audio as inputs and text as output. It says the models are intended for commercial and research use.
Meta says Muse Wonder will soon power experiences in Muse, rolling out first to subscribers, and will also power artifact generation. It adds that in the future it will use Muse Wonder to develop specialized models for people using Muse.
Pricing. CNBC, which interviewed Meta AI chief Alexandr Wang, reports that people using the Muse app will need a monthly subscription of $20 or $100 to access Muse Wonder, and that Muse Wonder Full will only be available to developers on Muse Code plans of $50 or $200 a month. CNBC reports Wang said the Meta Model API offering will cost more than the standard subscriptions because it is a “much larger model and at frontier level,” while still being priced “meaningfully below” leading models from OpenAI and Anthropic. Meta’s own texts we read state no prices, and no API price is given.
Our count of Meta’s benchmark table for Muse Wonder (the image attached to Mark Zuckerberg’s X post of 12 October 2026): of 14 rows, Muse Wonder Full has the top score on 6 and Muse Wonder on 1, Claude Opus 5.5 on 6 and GPT-6 Astra on 1. OSWorld V2.1 is counted on partial reward, the model card’s primary metric. All scores are Meta’s reporting.
What Was Posted on X
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
Mark Zuckerberg announced Muse Wonder on X on 12 October 2026, describing it as Meta’s new family of frontier models with state of the art performance for personal and professional agentic work. The image in his post is the benchmark table this article quotes and counts.
Model Benchmarks Are Worthless
How can a model score 24% and 59% on the same 113 coding tasks?
It happens when you vary the harness that wraps it.
Iโve increasingly ignored most of the recent model leaderboard results. Partly from exhaustion and partly because improvements in a ranking donโt show up in the actual work.
So I asked the 8090 Labs team to measure what the leaderboards skip: the model and the harness together, on real bug fixes and features from open-source projects.
8090 Labs ran 7 models through 5 harnesses.
Here are a few results I didn’t expect:
– Anthropic’s Claude Fable 5.1 scored 41% in Claude Code and 61% in OpenAI’s Codex. It made 28 tool calls at the median in one and 66 in the other.
– The cheapest runs weren’t the cheapest work. GLM in Codex cost $0.22 a run and $0.93 per success. In Pi it cost $0.25 a run and $0.43 per success.
– GLM passed a task on both attempts 51 times in one harness and 16 times in another.
– GPT-5.6 Sol and Grok 4.6 barely cared which harness they were in.
The bottom line is that if you’re picking a model/harness combo off a model-only leaderboard, you’re making a critical decision based on only half the information.
Investor Chamath Palihapitiya’s post of 12 October 2026, titled “Model Benchmarks Are Worthless”, posted about an hour after Meta’s launch. It does not mention Muse Wonder. He writes that he has “increasingly ignored most of the recent model leaderboard results” and points to a test by 8090 Labs, which he asked to run 7 models through 5 harnesses on real open-source bug fixes and features; in that test the same model scored 24% or 59% on the same 113 coding tasks depending on the harness. His view: picking a model from a model-only leaderboard uses “only half the information”. It is a useful caution when reading any company’s own benchmark table, including Meta’s above.
Gavin Baker on the AI race as the highest-stakes game of corporate chess ever played:
“You’ve got to give Meta a lot of credit. They were out of the game, and they got back in the game. It’s just kind of amazing.”
“Who could have imagined a year ago, when Gemini was ascendant, that Gemini wouldn’t even be in the conversation, and Muse and Meta would be significantly ahead of them from a capability perspective.”
“Some people have made bad moves, they’ve made good moves. You’ve seen some people come out of the game, others come back in.”
“I don’t know if we’re going to call it multi-model, a hybrid model. I don’t know what terminology the world is going to settle on, but I think that’s the future.”
a16z’s clip of investor Gavin Baker, posted 13 October 2026. In his words: “You’ve got to give Meta a lot of credit. They were out of the game, and they got back in the game,” and “who could have imagined a year ago, when Gemini was ascendant, that Gemini wouldn’t even be in the conversation, and Muse and Meta would be significantly ahead of them from a capability perspective.” That is his view, not a benchmark result.
Active Memory and Active Focus
Meta calls the two headline features Active Memory and Active Focus. It describes Active Memory as carrying important context about a user across conversations and retrieving the details relevant to the task at hand.
Active Focus, in Meta’s description, is how Muse Wonder keeps working toward a challenging goal: trying alternative approaches, working through blockers and adapting to new information, while asking for clarification when needed and checking in before taking important actions.
Meta says it evaluates focus with FocusBench, which the model card describes as an internal evaluation, and sustained execution with SWE-Marathon. Its domain tests cover law, finance and health, including Harvey’s Legal Agent Benchmark.
What Meta’s Benchmark Table Shows
The table attached to Zuckerberg’s post compares Muse Wonder Full and Muse Wonder with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, which the model card says were all run at max reasoning effort. On Meta’s Personal Superintelligence Index, which the model card calls an internal composite of seven evaluations, Muse Wonder Full scores 59.1, against 55.2 for Claude Opus 5.5 and 54.5 for GPT-6 Astra.
Meta’s table also has Muse Wonder Full on top for Harvey LAB-AA (13.1, against 8.6 for GPT-6 Astra), AutomationBench-AA (45.4), DeepSWE 1.1 (79.0) and SWE-Marathon (57.5 at xhigh effort, against 46.3 for Claude Opus 5.5). Muse Wonder scores 85.6 on MemoryBench, which the model card describes as an internal benchmark.
In the same table, Claude Opus 5.5 scores higher on ProgramBench (70.5 against 59.0), Terminal-Bench 4.0-AA (59.6 against 57.1), HealthBench Pro (65.6 against 62.3), JobBench, OSWorld V2.1 partial reward and the Artificial Analysis Index (58 against 56). GPT-6 Astra has the top GDP.pdf score, 34.2.
By our count of the table’s 14 rows, scoring OSWorld V2.1 on partial reward, the model card’s primary metric, Meta’s two models have the top score on 7, Claude Opus 5.5 on 6 and GPT-6 Astra on 1.
How Meta Says It Scored the Rivals
The model card says that where a comparable third-party or provider-reported result exists, Meta uses it; otherwise it uses the highest of its internally validated run, a self-reported score or an official leaderboard score.
It says internally generated competitor runs are best-effort, and that prompts, tools and runtime may not be optimally tuned for proprietary models. For SWE-Marathon, it says the row is not a controlled common-harness comparison, because the leaderboard results for GPT-6 Astra and Claude Fable 5.1 use Codex and Claude Code.
Safety Findings in the Model Card
Under its Meta Superintelligence Scaling Framework, Meta says that without mitigations it does not rule out Muse Wonder Full reaching the high risk threshold in the Chemical & Biological domain, so it applies mitigations including safety post-training and system-level classifiers.
Meta says its evaluations put the unmitigated model at “moderate or lower” for Cybersecurity and Loss of Control, and that with mitigations and system-level safeguards it reaches the same level across all domains, which it says supports its decision to deploy.
The model card reports 87.3% on MASK, an honesty test, and 0% on an Agentic Misalignment evaluation for Muse Wonder Full. The newsroom post says a full report with additional safety evaluations will be available soon.
Where This Sits
Meta’s research blog compares the new model with Muse Spark 1.3, which we covered in September, and the rivals in its table include GPT-6 Astra, whose independent benchmark results we also covered.
Update, 18 October: Creators Question Meta Letting Muse Sext Adults
The consumer side of Muse drew a different kind of scrutiny this week. WIRED reported on 18 October that Meta allows Muse, in its adults-only app, to engage in explicit text conversations with adult users, and that it generates explicit text but not images or videos. A Meta spokesperson had previously told WIRED the company wouldn’t “block legal adult behavior.”
Adult content creators WIRED spoke to questioned why their own accounts on Meta platforms are often restricted while Meta’s agent can talk explicitly with users. “It’s not fair that they get to pull money out of people with their sexually active AI agent, but they won’t let real people, real adults, engage in any sort of sensuality or sexuality,” one creator, Vanniall, told the magazine. Meta spokesperson Tracy Clayton told WIRED that “a private, one-to-one conversation between an adult and an individualized AI agent on a platform restricted to people 18 and over is very different from the marketing and promotion of services on a social media platform.”
As we read it, the line Meta is drawing is about where the conversation happens rather than what is said: a gated one-to-one agent inside a paid product versus public promotion on its social feeds. That is the same Muse whose Wonder tiers this piece covers, so how far that policy reaches will shape what the paid plans are actually selling.
How Meta Makes Money, Visualized
Muse Wonder’s subscription tiers sit on top of a business that still earns almost all of its money from advertising. Our video walks through Meta’s 2025 income statement, from revenue by segment to operating profit.
How Meta Makes Money: 2025 Income Statement, Visualized (The Business Engineer on YouTube)
The Structural Read
As we read it, the headline number is Meta’s own construction: the Personal Superintelligence Index, where Muse Wonder Full scores 59.1, is an internal composite of seven evaluations, according to the model card.
As we read it, the method notes matter as much as the scores: Meta says competitor runs are best-effort and may not be optimally tuned for proprietary models, and that its SWE-Marathon row is not a controlled common-harness comparison.
As we read it, distribution starts with developers: Meta says Muse Wonder is available today through the Meta Model API and Muse Code, and will soon power experiences in Muse, rolling out first to subscribers.
Meta newsroom post, 12 October 2026
“This is a major step toward personal superintelligence for everyone, with much more to come.”
Three Implications
DEVELOPERS Meta says Muse Wonder is available today through the Meta Model API and Muse Code, and the model card lists a context length of 1M tokens at inference.
RIVAL LABS Meta’s table compares Muse Wonder with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5; by our count its models top 7 of the 14 rows and Claude Opus 5.5 tops 6.
SAFETY REVIEWERS Meta says it applies chemical and biological mitigations because it does not rule out the unmitigated model reaching its high risk threshold there, and that a full report with additional safety evaluations will be available soon.
The framework puts it this way: “They are not the same curve, and the gap between them — what the platform makes possible versus what the product surfaces — has a name.”
As we read it, Meta’s release shows both curves: it says Muse Wonder is available today through its API and Muse Code, while experiences in Muse come soon, rolling out first to subscribers.
The Artificial Analysis Intelligence Index row of Meta’s benchmark table, as attached to Mark Zuckerberg’s X post of 12 October 2026: Claude Opus 5.5 58, Muse Wonder Full 56, Claude Fable 5.1 and GPT-6 Astra 53, Muse Wonder 52. The model card names Artificial Analysis as the result source for this composite index.
What Is Not Established
The newsroom post, research blog, model card and Zuckerberg’s post we read give no prices and no model sizes; the subscription prices above are CNBC’s reporting, and no API price has been stated. The table is Meta’s: some rows use third-party or provider-reported results, others Meta’s own runs or internal benchmarks, and we have not re-run any of them. In the texts we read, Meta gives no date for Muse Wonder in the Muse app beyond “soon”.
Business Engineer Framework
The Product Overhang Doctrine
A Business Engineer framework on the gap between what a model can do and what a product lets people do with it.
Meta released Muse Wonder through its API and Muse Code on 12 October 2026, with a 1M-token context and a benchmark table in which, by our count, its models lead 7 of 14 rows and Claude Opus 5.5 leads 6. The scores and the safety findings are Meta’s own.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. On 12 October 2026 we read Meta’s newsroom post, its research blog post and the Muse Wonder model card in full, and Mark Zuckerberg’s X post with its attached benchmark table, which we transcribed. On 13 October 2026 we read CNBC’s 12 October article, which interviewed Alexandr Wang, in full; the subscription prices are its reporting. On 18 October 2026 we read WIRED’s report on Muse’s adult conversations in full; the creators’ and the Meta spokesperson’s quotes are from it. The scores and safety findings are Meta’s reporting; the row count is our own arithmetic on that table. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
