Meta Muse: partners give no figures, Instacart ads bypassed
- Bad news: Muse orders on Instacart do not support its ads, about 30% of revenue, so ad-reliant partners may restrict Muse.
- Bad news: Booking and Shopify mentioned Muse on 3 Nov calls but gave no Muse figures; no partner has shown real transaction volume.
- Mixed: Bug bounties paid out about $1.2 million, a good sign that also shows a real attack surface.
Read the full notes
Neutral to slightly negative today. The Muse product experience is getting stronger (it will draw an Instacart cart itself, keeping users inside Muse), but the economics for merchants are weaker.
Read more
Instacart's third-quarter advertising and other revenue was $331 million (about 30% of revenue, up 23% year on year), and those ads are not supported in Muse orders. Amazon blocked Muse's core use for just this reason, since on-site ads get bypassed, so the more a partner depends on ads, the more likely it is to restrict Muse. Booking (OpenTable as the first dining partner, Agoda for the Asia launch) and Shopify mentioned Muse on their 3 Nov earnings calls, but neither gave any Muse figures. Since DoorDash's 'a few hundred orders', no partner has offered real evidence of transaction volume. Section 15 of Wells Fargo's online banking agreement, effective 24 Nov, treats instructions from AI agents as authorised by the user and puts the risk on the user. After the blocks by JPMorgan and Schwab, banks have changed tack, but Muse handling financial tasks will struggle to scale in the short term. On security, the $1.2 million in bug bounties (up to $300,000 for a single authentication flaw and $90,000 for a prompt injection) is a good sign but also shows a real attack surface. The model's AA Intelligence Index score of 50 trails Gemini 4 Carbon at 59, Claude Opus 5.5 at 58 and GPT-6 Astra at 57 by about 9 points. On the AA-Briefcase knowledge-work benchmark, though, its 1659 is behind only Kimi K3.1 Pro at 1678. On Endor Labs' agent security leaderboard, functional correctness of 91.6% is the best of all, but a security pass rate of 25.1% ranks only seventh, and total cost of $692 is expensive.
Key evidence
- Juozas Kaziukėnas, founder of Marketplace Pulse, said on X on 3 Nov that Muse will launch a native, connector-driven shopping cart, with Instacart first (not officially confirmed by Meta)
- Instacart's help centre (cited by FourWeekMBA) lists what Muse orders do not support: embedded checkout, membership and promotional prices, coupons, Instacart ads, pickup, substitution preferences and browsing by aisle
- Instacart third-quarter 8-K (29 Oct, checked against the SEC filing): revenue $1.095 billion, advertising and other revenue $331 million (about 30%, up 23% year on year)
- Booking earnings call, 3 Nov (Benzinga transcript): OpenTable has just become Muse's first dining partner and Agoda is its launch partner in Asia. Management quantified only the OpenAI channel (one of its fastest-growing channels) and gave no Muse data
- Shopify, 3 Nov (CNBC): Finkelstein said agents have not pushed prices down, and that this is the first holiday season in which AI-assisted shopping goes mainstream. It did not disclose orders driven by Muse
- muse.ai/blog/securing-muse (2 Nov, David Singleton): the bug bounty ran privately from April and publicly from 8 Sep, receiving 630+ reports, of which 64 were valid, paying out about $1.2 million in total
Show 5 more
- The largest single payout was $300,000, for an authentication flaw in the four months before launch that could read conversations and tokens on internal-test virtual machines. A $90,000 prompt injection attack (disguised as Muse browsing-tool status messages) could make the agent carry out actions that should have been blocked
- Tech Times quotes: Wells Fargo updated Section 15 of its online banking agreement on 2 Nov, treating instructions from any AI agent acting with account credentials as authorised by the user, with no further verification and at the user's own risk, effective 24 Nov (original not checked directly)
- Artificial Analysis Intelligence Index v5 (Tom's Hardware and others): Gemini 4 Carbon 59, Claude Opus 5.5 58, GPT-6 Astra 57, Muse Wonder Full 50, a gap of about 9 points
- AA-Briefcase knowledge-work benchmark: Muse Wonder 1659, above Gemini 4 Carbon at 1609 and Grok 4.7 at 1638, and behind only Kimi K3.1 Pro at 1678
- Endor Labs agent security leaderboard, 3 Nov: Muse Code + Wonder has functional correctness of 91.6%, first overall, and a security pass rate of 25.1%, only seventh. Total cost is $692, which is expensive
Catalysts to watch
- Instacart native cart goes live (officially confirmed by Meta)
- Any partner (Booking, Shopify, Instacart, DoorDash) discloses specific Muse order volumes
- Wells Fargo agreement takes effect on 24 Nov: watch whether other banks follow with the same liability shift
- Whether the next Muse model upgrade narrows the 9-point gap to Gemini 4 Carbon, Claude Opus 5.5 and GPT-6 Astra
- Endor Labs security pass rate improves (from 25.1%)
Risks
- Merchants' advertising interests conflict structurally: the more a partner depends on ads, the more likely it is to restrict Muse
- Banks shifting liability could become standard practice, holding back Muse in financial use cases
- Prompt injection has been shown to make the agent carry out actions that should have been blocked, and the attack surface grows once it connects to email and payment accounts
- Overall intelligence trails the frontier models by about 9 points
- Total cost of $692 is expensive, which affects commercial pricing for agents
- No partner has yet disclosed real Muse transaction volumes
Sources 10
- Marketplace Pulse (Juozas Kaziukėnas on X, 3 Nov, to be confirmed)
- Instacart help centre (cited by FourWeekMBA)
- Instacart 8-K, 29 Oct (SEC)
- Booking earnings call, 3 Nov (Benzinga)
- Shopify, 3 Nov (CNBC)
- muse.ai/blog/securing-muse, 2 Nov
- Tech Times, 2 Nov (Wells Fargo agreement)
- Artificial Analysis Intelligence Index v5
- AA-Briefcase knowledge-work benchmark
- Endor Labs agent security leaderboard, 3 Nov