Mark Zuckerberg had not posted on X in about three years. On Thursday he broke that silence to announce Muse Spark 1.1, Meta's new AI model built for agentic tasks and coding, with one claim front and center: it is strong, and it is cheap.
What Muse Spark is
Muse Spark 1.1 comes from Meta Superintelligence Labs, the group Meta stood up after a rocky stretch in the AI race. It is natively multimodal, meaning it handles text, images, and video, with a memory of up to a million tokens, enough to hold long documents and multi-step projects in mind at once. It is built for work where the model doesn't just answer but acts: using tools, writing and running code, and even operating a computer. In Meta's demos it clicked through apps like Facebook Marketplace on its own.
The pitch: near the top, at a fraction of the price
Meta says Muse Spark rivals the best models from OpenAI and Anthropic on agentic tasks. Meta's Alexandr Wang called it "industry-competitive," and the developer pricing undercuts the frontier badly: about $1.25 per million tokens of input and $4.25 for output, several times cheaper than the going rate for GPT-5.5 or Claude Opus 4.8. New accounts get $20 in free credits.
Early hands-on reactions were mostly positive. One user had it one-shot a full presentation on the history of U.S. men's soccer, complete with researched facts and gathered images. Another built a clean front-end web page from a single prompt. A Meta researcher summed it up as "a quite strong model with a very good price" focused on agents and coding.
Two ways in
For most people, Muse Spark shows up inside Meta AI, the assistant already built into the Meta AI app, Instagram, WhatsApp, and Messenger, at no cost. For developers, Meta opened a new Meta Model API, its first real move to sell access to its own models the way OpenAI and Anthropic do. That API is the bigger strategic story: Meta spent years giving models away as open weights, and now it wants a seat at the paid-API table too.
We covered Meta's Muse Image, its first in-house image model, just a few days ago. Now Muse Spark is the coding and agent counterpart, and it undercuts the top labs on price. For a company that has looked a step behind in AI, that is two serious launches in a week.
The real winner is the consumer. With OpenAI, Anthropic, Google, xAI's Grok, and now Meta all shipping competitive models, the pressure on price is constant, and it lands in our favor: more capability for less money, and in Meta's case, free inside the apps you already open every day.
I want to put this one to the test myself, especially the coding and computer-use side, and see whether it holds up against Fable 5 and GPT-5.6 in real work or just on the benchmark charts.
Sources
- Mark Zuckerberg (@finkd)
- Alexandr Wang (@alexandr_wang)
- Meta for Developers (@MetaforDevs)
- Meta, "Introducing Muse Spark" (Meta Superintelligence Labs)
- CNBC, "Meta jumps into AI coding market to chase Anthropic and OpenAI"
- Axios, "Meta updates its Spark model, releases developer version"
- Demo: @tjshibata one-shot presentation