#摩根大通将Meta目标价上调至820 JPMorgan: Meta’s models and agents are in place, upgraded to a moderately positive view!
Met
The relevant market price has rebounded 20% from its recent low, but is still down 1% year to date, while the S&P 500 has risen 12% over the same period. The core judgment of this JPM report is that Meta’s in-house model team has largely delivered on its goal of reaching the frontier within a year. From July’s Muse Spark 1.1 to the recent 1.3, it can now rank alongside the Claude and GPT series on benchmarks covering agents, coding, instruction following, and long-context capabilities. This model capability is the prerequisite for Meta to open up revenue streams from agents, enterprise tools, and developer interfaces beyond advertising.
First, the models themselves. Muse Spark 1.3 is the latest version, and the report believes its competitiveness comes from a rapid iteration cycle, attractive pricing, and continuously expanding computing power. The next-generation Watermelon model has undergone more advanced pretraining and is expected to launch soon, while models following Watermelon are already being trained at scale on the Prometheus GW cluster in Ohio. Whether model quality can continue to keep pace will determine the ceiling for all subsequent monetization paths. 1. Muse agent reaches third place in the US App Store free rankings two days after launchMuse is Meta’s consumer-focused personal agent and has launched for US users on iOS, Android, and the web. It uses its own virtual computer to browse websites and operate interfaces, helping users complete tasks such as online shopping, ordering food, filling out forms, and communicating via email and text messages. Its training covers platforms including Instagram, WhatsApp, Spotify, DoorDash, Etsy, Reddit, Yelp, and Outlook/Gmail. Early data provides two reference points: Muse usage reached 10 times the size of the test queue, and it rose to third place in the US Apple App Store’s free app download rankings on the second day after launch. Muse is currently free and provides 100 million tokens per week. The paid tiers offer 500 million tokens per month for $20 per month and 3 billion tokens per month for $100 per month. The report believes Meta’s current focus is not monetization, but building awareness, adoption, and usage habits; it is more likely to adopt a transaction-fee or commission model in the future.
Reaching third place in the free rankings on the second day shows that Meta’s distribution capabilities can indeed quickly put a product in front of users, but download rankings and retention are two different things. I am more focused on whether usage 10 times the size of the test queue can persist after the free quota is exhausted, as this will determine whether the commission model has a real transaction base.
2. Enterprise agents charge by token, with over 1 million businesses already using them weekly
Meta Business Agent is a plug-and-play tool used via subscription. It can answer questions, mention products, schedule appointments, and qualify leads. As of Q2 2026, more than 1 million businesses were using the agent weekly on WhatsApp and Messenger. Business Agent Platform is an API-based enterprise layer that can connect to systems such as Shopify, Zendesk, and Shopee, allowing agents to perform actions on behalf of businesses. The billing model has already been implemented: starting August 1, API and enterprise users will be charged by token, at $2 per 1 million tokens. A single message consumes 20,000 to 25,000 tokens, equivalent to approximately 4 to 5 cents per message. Management also mentioned that enterprise-agent monetization could eventually evolve into an outcome-oriented bidding system similar to the advertising business. The potential of this segment lies in Meta’s existing relationships with millions of advertisers and small and medium-sized businesses, allowing agent adoption to reuse this distribution channel.
3. Model interfaces and coding agents compete for developer share with low pricing$META