Meta just released a new AI model called Muse Spark, and it does not just chat, it can hop between apps and finish tasks on your behalf. If you already use ChatGPT for daily questions, this is the kind of release that shows how the next wave of AI is starting to feel less like a chat and more like a helper you can hand things off to.
The Gist
- Meta released Muse Spark on July 9, 2026, its most advanced AI model to date.
- Muse Spark is agentic, meaning it can plan several steps and click through apps to finish a task.
- It is available in preview inside the Meta AI app and via the Meta Model API.
- Its price sits well below the top rivals for the same tier of quality.
Have ChatGPT Recap This Article
ChatGPTMeet Muse Spark, Meta’s own answer to ChatGPT
If ChatGPT is the AI most people met first, Muse Spark is Meta’s version of the same idea, only pushed a step further. It was built inside a team called Meta Superintelligence Labs, which is basically the top research arm at Meta focused on building the strongest AI models the company can ship.
The version released on July 9 is Muse Spark 1.1. You can already use it inside the Meta AI app when you turn on the Thinking mode, and developers can call it directly through an API. This is Meta finally putting a serious chat rival right next to ChatGPT, Claude and Gemini in everyday tools.
On the same day, one thing became obvious: Meta is not trying to be the smartest, it is trying to be the cheapest at the top table. That matters because AI models are getting close to each other on raw intelligence tests, and price is what will decide which one ends up inside the apps you already use.
If you were curious about how Meta stacks up against Anthropic and OpenAI on the money side, the way Anthropic just overtook ChatGPT’s maker on revenue is a useful anchor. Meta is joining that same top tier with Muse Spark.
How Muse Spark uses your apps for you
Meta describes Muse Spark as a model that reads images, videos and PDF files, and that can pilot other software while you watch. The idea is that you give it a goal, and it clicks, scrolls, fills in and moves to the next tool on its own to reach the end of the task.
Under the hood, Muse Spark can call in smaller helper agents. Think of a general contractor who calls a plumber for one task and an electrician for another. The main model plans, the smaller ones execute pieces in parallel, and everyone reports back. That is the mechanic Meta is pushing as its edge.
The model can also remember a lot at once. Meta gave it a working memory of one million tokens, which is roughly the size of a very long book kept open at all times. That means Muse Spark can hold the whole story of a big project inside its head, instead of losing the thread after a few messages.
If this feels similar to the way Gemini Spark now does your tasks on a Mac, that is the right instinct. Both are betting on a future where the AI is not just something you type into, it is something that clicks the interface with you.
Keep learning on AI Noobies:
- Claude Reflect Shows How You Actually Use AI
- 5 Common ChatGPT Mistakes Beginners Make Daily
- GPT-5.6 Rolls Out Worldwide Starting Today
What “agentic” and “tokens” actually mean
Two words come up a lot around Muse Spark, and they are worth pinning down. The first one is often called an agentic AI. It just means the model can act, not only answer. It can open a tab, run a search, download a file, put together a form, and then decide the next step by itself. It is closer to hiring an intern than typing into a chat.
The second is the count of tokens. Tokens are how AI companies count what you send and what the model writes back. One short paragraph is roughly a hundred tokens. Meta priced Muse Spark at 1,25 dollar per million tokens you send, and 4,25 dollars per million tokens it writes back. Very cheap by frontier standards.
The nuance here is that Muse Spark is not the best model out there. Independent testers put it neck and neck with the top Chinese model GLM-5.2, both a step behind Claude Fable 5. The reason it still matters is the cost curve. When you run agents in the real world, they use a lot of tokens over hours, and the cheaper option ends up being the one companies pick.
Zooming out, this is the kind of shift that changes what changes for you. In the wider public debate, expect to hear more about which AI is quiet and cheap enough to slip into your banking app, your travel app or your school app, without you noticing the switch.
How to try Muse Spark this week
Getting your hands on Muse Spark is simpler than it sounds. Open the Meta AI app or go to meta.ai in your browser. Look for the toggle called Thinking mode. Once it is on, you are chatting with Muse Spark. Ask it to help plan a trip, summarize a long PDF you upload, or extract the numbers from a screenshot of your bill.
A good first test is to give it a task that has three steps. Something like: read this two-page rental contract, translate the risky clauses into simple English, and then draft a polite email to the landlord asking for two changes. That is exactly the kind of chain a plain chat used to fumble and that an agentic model handles cleanly.
If you want to compare Meta’s new model with the alternatives before picking your daily driver, our side-by-side of Claude, ChatGPT and Gemini for real beginner use cases stays a solid map. Muse Spark now joins that same conversation as a fourth serious option.
The takeaway is not that you must drop your current AI and switch to Muse Spark today. The takeaway is that the market is now four horses deep at the top, and the one that hides inside the app you already use may end up being the one you use most, without ever making a choice.
Stay tuned on AI Noobies.
About the Author
1 thought on “Muse Spark: Meta’s New AI Model That Uses Your Apps”