Meta Muse ranks 4th on GDPx, first in cybersecurity

SPCXTuesday, 3 Nov 2026

Mixed: Muse Wonder Full is 4th on GDPx at 9.0% and first in cybersecurity, but its low refusal rate is a safety and regulatory risk.
Bad news: Zuckerberg called Muse very compute-intensive and not yet profitable, and Meta's Q3 free cash flow was -$2.12 billion.
Bad news: OpenAI's self-graded test has GPT-6 Astra at 80%, and CNN names Meta as its biggest rival, so Muse's first-mover window is narrowing.

What we watchWhether Meta delivers its committed 10x or greater cost-efficiency gain within a year.

Read the full notes

Three signals converge. (1) Muse Wonder Full is in the top tier on SemiAnalysis's GDPx: 4th at 9.0%, 3pp behind Claude Opus 5.5 at 12%. It is number one in cybersecurity by a clear margin, partly because of its low refusal rate. For an agent that reads users' emails and operates their browsers, that is a safety and regulatory risk, not a pure positive.

(2) Industry economics are breaking down. On 1 Nov Musk admitted that Grok Bot's cost against its revenue is unsustainable, and xAI is cutting weekly quotas despite 800,000 weekly paying users, per the SpaceX investor update. On the Q3 call Zuckerberg called Muse very compute-intensive and unprofitable, and Meta's Q3 FCF was -$2.12 billion per the 8-K. The question is shifting from whether it can work to what it costs, and pricing (free plus a $20-100 monthly subscription) may not cover unit economics. (3) OpenAI benchmarked GPT-6 Astra itself at 80% on daily tasks against 49.3% for GPT-5.2 (self-graded, so discount accordingly), and a CNN report names Meta as OpenAI's biggest competitor in personal agents, so Muse's first-mover window is narrowing. Net: the model capability box is ticked, but the promise of cost efficiency (Meta has committed to 10x or more in a year) is now the critical variable. Meta's ad cash flow lets it keep free tiers open while cash-constrained competitors such as xAI impose limits, a potential window for user acquisition.

Key evidence

SemiAnalysis GDPx benchmark, 2 Nov: 206 tasks across 11 domains (finance, legal, SWE, cybersecurity and others). The numbers are consistent across the official tweet and multiple outlets
Leaderboard: Claude Opus 5.5 12.0%, Gemini 4 Carbon 10.9%, GPT-6 Astra 9.6%, Muse Wonder Full 9.0% (4th), Gemini 4 Argon 7.3%. Standard Muse Wonder is below Argon
SemiAnalysis, direct quote: Muse Wonder Full is the best cybersecurity model currently, by a clear margin, partly because of its low refusal rate (Claude and GPT are held back by safety classifiers)
Wonder Full has no public pricing, so it is excluded from the cost-efficiency rankings
Caveat: a low refusal rate in an agent that operates on the web and reads email is a safety and regulatory risk, not a pure positive
Musk's X post of 1 Nov (Economic Times, 2 Nov): Grok Bot users are exhausting their weekly quota early. Musk's reply: the cost of serving Grok Bot was unsustainably high relative to revenue

Show 8 more

SpaceX investor update last month: Grok Bot has 800,000 weekly paying users
Zuckerberg on the Q3 earnings call: Muse is very compute-intensive and not yet profitable
Meta's Q3 FCF was -$2.12 billion per the SEC 8-K. The free-tier quota is not disclosed
Meta has committed to a 10x or greater cost-efficiency improvement in one year, which is the critical variable
OpenAI daily-task benchmark via CNN, 2 Nov: a test designed from a survey of about 4,000 users (find a kindergarten, a diabetes meal plan, forms, a shopping cart). GPT-6 Astra scored 80% overall and GPT-5.2 scored 49.3%
The CNN report names Meta as OpenAI's biggest competitor in personal agents, but this is OpenAI benchmarking itself, so discount accordingly
Muse's first-mover window is narrowing as ChatGPT frames everyday task execution as the main battleground
No new official Muse user numbers today

Catalysts to watch

Can Meta deliver its committed 10x or greater cost-efficiency improvement within a year? That decides whether the free tier stays open
Will Meta announce Muse quota limits, following the xAI Grok pattern? That would signal that unit economics are still broken
Independent reproduction of OpenAI's daily-task benchmark by third parties
The next SemiAnalysis GDPx update: can Muse close the 3pp gap to Claude Opus 5.5, or is it structural?
Any regulatory response to the low refusal rate in agents with email and web access
Muse WAU/MAU disclosure in the Q4 2026 earnings or earlier

Risks

The low-refusal advantage is a safety and regulatory liability. It could trigger compliance action or forced constraints that erase the cybersecurity number one
The pricing model (free plus a $20-100 subscription) may not cover compute cost, repeating the xAI pattern
Meta's FCF of -$2.12 billion shows the cost pressure is real, not theoretical
OpenAI framing Meta as its main rival accelerates competitive dynamics, and the first-mover window is shrinking
Every industry participant is losing money per task, so the market may go through painful price discovery

Show 2 more

Model capability is no longer a differentiator (top tier is table stakes). Distribution and cost become decisive: Meta has distribution through Facebook and Instagram, but cost is still an open question
Independent benchmarks will not resolve consumer behaviour. Actual user data is needed (DAU, retention, task completion)

Sources 7

SemiAnalysis GDPx benchmark, 2 Nov (official tweet and multiple media)
Elon Musk X post, 1 Nov
Economic Times, 2 Nov (Grok Bot cost comment)
SpaceX investor update (Grok Bot 800,000 weekly paying users)
Meta Q3 2026 earnings call (Zuckerberg on Muse)
SEC 8-K, Meta Q3 2026 (FCF of -$2.12 billion)
CNN, 2 Nov (OpenAI daily-task benchmark and Meta framing)

A brief is a lead for our research, not a verified conclusion or advice.