Meta is making a lower-cost play for teams building AI coding agents.
The company released Muse Spark 1.1 in US public preview on Thursday, opening access to developers building coding assistants that can use tools, inspect files, and work through longer software tasks. Pricing starts at $1.25 per million input tokens and $4.25 per million output tokens, giving teams a lower-cost option as they compare agentic coding models from Meta, OpenAI, Anthropic, and others.
For engineering teams, the pitch is simple: more room to experiment. The harder question is whether cheaper agents can be trusted near real code without tighter review, logging, and approval rules.
‘Strongest at agent performance, tool use, and computer use’
US users who join the preview receive $20 in free credits before paid usage begins.
Early access had been limited to select partners. Public preview opens testing to US developers working on coding assistants, app development, and agent-based workflows. OpenAI-style compatibility may also reduce setup work for teams comparing Muse Spark 1.1 with other AI models.
Mark Zuckerberg called the release “a strong agentic and coding model at a very low price” in a post on X, adding that it is strongest at agent performance, tool use, and computer use.
Alexandr Wang, Meta’s chief AI officer, told CNBC the pricing is “very aggressive and attractive” and said the company wants pricing that “scales with immense consumption usage.”
The model can work across files and screens
Meta says Muse Spark 1.1 is designed for several kinds of work:
- Writing and debugging code
- Using software tools
- Understanding text, images, video, and documents
- Completing multi-step tasks with less human direction
One major selling point is its ability to work across different kinds of inputs. A developer could ask the model to reason through written instructions, review a screenshot, read project files, or operate inside a software environment as part of the same job.
The company also says the model supports a 1 million-token context window, which could help it keep track of longer projects or large sets of instructions.
Why coding is a key test for Muse Spark 1.1
Software development is a tough test for Muse Spark 1.1 because generated code still has to work after the answer is written. A failed test may force the model to trace the error before touching the source code. Once it makes a change, the fix still needs to hold.
Visual bugs create a different problem. Not every failure appears in a test result. A broken page may require the system to read a screenshot before it opens a sandboxed browser or reviews the project files.
Teams still need guardrails before real code
Lower pricing could attract more startups and engineering teams to AI-assisted coding tests, particularly if they have avoided running agents at scale due to cost.
Access alone does not make automation safe, though. A coding agent may speed up routine work, then create a new problem by changing the wrong file or missing context that a human reviewer would catch.
Before using these systems on live code, engineering leaders need rules for where the tool can operate and when a person must approve the next step. Logs also become important because teams need to know what changed if an AI-assisted fix breaks later.
Cheaper testing may help developers move faster, but responsibility will stay with the team using the tool. If an agent introduces a bug into a product or internal system, customers and employees will deal with the fallout first.
Meta is facing renewed pressure in France over whether news publishers should be paid for content that supports its platforms.