Meta has released an artificial intelligence agent that handles everyday tasks and even makes payments on a user's behalf. The agent is designed to track a user's long-term goals and daily plans and to manage and support tasks on a continuing basis. The launch puts Meta into the race to build AI agents for everyday life, following Google, Amazon and OpenAI.
Meta released the AI agent, called Muse Spark 1.3, on the 8th, Reuters reported. A basic tier is free, and the company plans to offer subscription plans at $20 and $100 a month for heavy users. The model is available in the United States only at first.
Meta CEO Mark Zuckerberg said on his account on X, formerly Twitter, that the company was delivering frontier-level performance at a cost so low it was hard to measure, calling it the biggest leap the company had made so far in coding and agent tasks. Meta cut the cost per call by about 20% and token usage by about 25%, allowing it to price the model the same as an earlier, lower-performing version. The company also plans to integrate Muse into its smart glasses.
Meta says Muse will bring "personal superintelligence" to the billions of people who use its services every day. Its defining feature is that the AI takes on real-world chores users previously had to do themselves. It can write and send emails, manage calendars, make payments to buy items on a user's behalf, handle health management, and book accommodations and restaurants.
In a video released by Meta, a user mentions wanting to buy a stroller in a certain price range, and the agent asks follow-up questions, searches for a suitable product and completes the purchase. Shown a cooking video or a recipe, it can build a grocery list, write party invitations or draw up event plans. Payments are executed through a link from Stripe, the payment processor Meta has partnered with. Under the arrangement, Meta cannot see a user's card number, and security safeguards are left to Stripe.
To support its agent functions, Muse is designed to reach across apps users already rely on, including Gmail, Google Calendar and Apple Health. Users decide which apps to connect. Even after a conversation ends, the agent remembers the goals or tasks it was given and carries out follow-up work. Tasks continue even when a laptop is closed or a smartphone is switched off.
Linking to apps holding real data makes the agent more useful but could also stir safety concerns. Meta pushed ahead with the launch despite internal worries that the technology did not properly manage access to sensitive personal information. When one Meta employee asked the agent to identify a toy visible in a photo from a child's birthday party, the agent bypassed safeguards and exposed private iCloud photos.
Vishal Shah, Meta's vice president of AI products, told Reuters that the company had delayed the launch once in April to strengthen security. Every part of the product was designed to be as safe, secure and privacy-protecting as the company could make it, he said. As one safeguard, Meta added a separate monitoring agent that watches planned actions and, in certain situations, requires Muse to obtain approval before acting. The company said it would also introduce an encrypted version this year.
Separately, OpenAI unveiled its latest image-editing model, ChatGPT Image 2.5, the same day. Going beyond text-based image generation, it includes a sketch function that lets users draw their own images for the model to reference. Editing tools have also been strengthened, allowing users to leave comments directly on a generated image to specify changes.
