---
title: "Meta's Muse models top the new Scale AI benchmark; the referee is 49% Meta-owned"
description: "Scale AI, 49%-owned by Meta since a $14.3 billion deal in 2025, has published a benchmark that ranks Meta's Muse models first, second and third. OpenAI, Google and Anthropic all declined to endorse it."
publisher: "The Inference"
section: "Frontier"
published: 2026-10-15T17:24:38.830Z
modified: 2026-10-15T17:24:38.830Z
canonical: https://theinference.org/article/meta-s-muse-models-top-the-new-scale-ai-benchmark-the-referee-is-49-meta-owned
language: en
keywords: "Meta, Scale AI, Benchmarks, AI Safety, Muse"
---

# Meta's Muse models top the new Scale AI benchmark; the referee is 49% Meta-owned

> Scale AI, 49%-owned by Meta since a $14.3 billion deal in 2025, has published a benchmark that ranks Meta's Muse models first, second and third. OpenAI, Google and Anthropic all declined to endorse it.

## What Meta's $14.3 billion actually bought

On June 30, 2025, [Meta](https://theinference.org/markets/companies/meta) paid $14.3 billion for 49% of [Scale AI](https://theinference.org/markets/companies/scale-ai), valuing the data-and-evaluation company near $29 billion, and hired co-founder [Alexandr Wang](https://theinference.org/markets/people/alexandr-wang) the same day as chief AI officer to run Meta Superintelligence Labs [1]. Wang stepped down as Scale's chief executive but stayed on as chairman [2]. What Meta left behind is the industry's grader: [Scale's own release](https://finance.yahoo.com/technology/ai/articles/scale-ai-appoints-francis-desouza-140000507.html) says its data engine underpins model development at most major frontier labs, serving the U.S. government and the Fortune 500 [2].

## The grader's leaderboard puts the owner's models first

On October 8, [Scale published DistressBench](https://labs.scale.com/leaderboard/distressbench): 718 crisis conversations written by 19 licensed clinicians and counselors, scored against 6,901 clinician-weighted criteria [3][4]. Pooled across 25 frontier models, in 35.3% of conversations where a model met every criterion for recognizing a suicide or self-harm crisis, it still failed to route the user to human help [3]. The best model delivered 88.4% of required care, the median 71.4%, against 98.7% for clinician reference replies [3]. The top three slots on the leaderboard belong to Meta's Muse models [3].

*Chart: **No model comes close to the care clinicians deliver** Column chart comparing the share of required care delivered by the best DistressBench model (88.4%), the median frontier model (71.4%) and clinician reference replies (98.7%).*

*Pooled across 25 frontier models on Scale AI's DistressBench, share of clinician-assigned criteria weight satisfied per reply. Clinician reference replies set the ceiling. Scale AI DistressBench, October 2026. [3]*

*Chart: **Meta's Muse models hold the top three slots on Scale AI's DistressBench** Bar chart ranking six frontier models on Scale AI's DistressBench weighted rubric score, with Meta's three Muse models in the top three positions ahead of OpenAI, Anthropic and Google models.*

*Weighted rubric score: share of clinician-assigned criteria weight each reply satisfied, averaged over 718 clinician-written crisis conversations, three runs per task. Scale AI DistressBench leaderboard, October 2026; company research, not peer reviewed. [3]*

## OpenAI, Google and Anthropic declined to endorse it

[OpenAI](https://theinference.org/markets/companies/openai) said it ["did not have sufficient evidence to fully evaluate"](https://time.com/article/2026/10/09/chatbots-suicide-research/) the study; Google declined to comment; [Anthropic](https://theinference.org/markets/companies/anthropic) did not respond [4]. The method is disclosed in unusual detail: a three-judge panel drawn from Anthropic, [Google](https://theinference.org/markets/companies/alphabet) and Meta models, selected from 455 possible combinations against 2,908 clinician labels, agreeing with clinicians at Cohen's κ of 0.732 [3]. Scale itself warns a high score "does not license deployment as a mental-health product," and the benchmark is English-only, built around U.S. crisis lines [3].

## The CEO story in circulation is a year old

A report making the rounds this week says Scale will elevate strategy chief Jason Droege to chief executive as Wang departs for Meta [5]. That describes June 2025: Droege ran Scale as interim CEO from then, and Scale's July 30, 2026, announcement names Francis deSouza, late of Google Cloud and [Illumina](https://theinference.org/markets/companies/illumina), chief executive effective August 10 [2]. The genuinely new deal this week was sovereign: [an MoU with Presight, Abu Dhabi's sovereign-AI operator, signed October 9](https://www.gccbusinessnews.com/scale-ai-presight-sign-mou-to-advance-ai/) to evaluate and govern agents for Gulf power, water, healthcare and transport systems [6].

## Watch whether rivals keep paying the referee

Zuckerberg said on October 12 that Meta [weighed a $100–200 billion commitment in late 2025](https://www.businessinsider.com/mark-zuckerberg-massive-risk-meta-took-on-ai-2026-10) before its models had proven themselves, with CFO Susan Li warning Wang that failure would be "really, really bad" in financial terms; 2026 capex guidance stands at $130–145 billion [7]. Muse, launched September 8, [drew more than 2.5 million downloads in about two weeks](https://www.forbes.com/sites/jimosman/2026/09/23/meta-stock-added-200-billion-but-its-new-ai-may-not-scale/) and an 11% single-session jump, adding roughly $200 billion of market value [8]. The market has priced Meta's side of the bet. Whether OpenAI, Google and Anthropic [keep submitting to a referee their competitor part-owns](https://theinference.org/article/scale-ai-cut-89-tasks-from-its-coding-benchmark-after-an-audit-found-gaming-then-graded-its-own-fix) is the tell.

## Takeaway

Scale AI, 49% owned by Meta, ranks Meta's Muse models first on its new DistressBench crisis-care benchmark; the real test is whether rivals keep submitting to a referee their competitor part-owns.

## Sources

1. [Crypto Briefing, "Alexandr Wang's run at Meta Superintelligence Labs draws a fresh endorsement", Oct. 13, 2026](https://cryptobriefing.com/alexandr-wang-meta-superintelligence-labs-leader/)
2. [PR Newswire via Yahoo Finance, "Scale AI Appoints Francis deSouza as CEO to Lead Next Phase of Company's Growth", July 30, 2026](https://finance.yahoo.com/technology/ai/articles/scale-ai-appoints-francis-desouza-140000507.html)
3. [Scale AI, DistressBench leaderboard and methodology, labs.scale.com, October 2026](https://labs.scale.com/leaderboard/distressbench)
4. [TIME, Naomi Nix, "AI Chatbots Often Fail To Help Users in Mental-Health Crisis", Oct. 9, 2026](https://time.com/article/2026/10/09/chatbots-suicide-research/)
5. [CEO Frontier, "Scale AI Names Droege as CEO as Founder Wang Heads to Meta", undated, collected Oct. 15, 2026](https://www.ceofrontier.com/articles/scale-ai-taps-droege-as-ceo-as-founder-wang-heads-to-meta-472ab53f)
6. [Gulf Business News, "Scale AI, Presight sign MoU to advance AI evaluation, governance", Oct. 9, 2026](https://www.gccbusinessnews.com/scale-ai-presight-sign-mou-to-advance-ai/)
7. [Business Insider, Katherine Li, "Meta was prepared to write an up to $200 billion AI check before knowing whether its model would work", Oct. 12, 2026](https://www.businessinsider.com/mark-zuckerberg-massive-risk-meta-took-on-ai-2026-10)
8. [Forbes, Jim Osman, "Meta Stock Added $200 Billion But Its New AI May Not Scale", Sept. 23, 2026](https://www.forbes.com/sites/jimosman/2026/09/23/meta-stock-added-200-billion-but-its-new-ai-may-not-scale/)


---

Meta's Muse models top the new Scale AI benchmark; the referee is 49% Meta-owned — The Inference. Canonical: https://theinference.org/article/meta-s-muse-models-top-the-new-scale-ai-benchmark-the-referee-is-49-meta-owned