Meta says Muse Spark 1.3 uses fewer tool calls and fewer tokens than the earlier version. Users should check if it also needs less help when a task goes wrong.
New to this? Read it in simple words
- Meta released Muse Spark 1.3 on September 2. Meta says it can work on longer AI agent tasks.
- Meta’s own engineers say it used about 20 percent fewer tool calls than version 1.2. They say it also used 25 percent fewer tokens.
- Fewer steps can save time and money. They can also mean fewer possible errors.
- Independent tests have not confirmed these numbers. Outside experts still need to test Meta’s safety claims.
- AI agent
- An AI that takes steps on its own to finish a task, such as booking or coding.
- Token
- A small piece of text, often part of a word. AI use is priced per token.
- Tool call
- When an AI uses another program, such as web search, to do part of a task.
Meta wants better work habits
Meta released Muse Spark 1.3 on September 2 through Muse Code and its model API. The company says the model can follow detailed instructions for longer. It can also manage several jobs in one conversation without losing the main goal.
Meta says the agent can notice missing information, ask a question, and pause before an important action. These habits are valuable because a safe agent should show doubt before it makes a change that is hard to undo.
Meta reports about 20 percent fewer tool calls and 25 percent fewer tokens than version 1.2. These are Meta’s own measurements.
Fewer steps can save time and money
Meta’s own engineers say Spark 1.3 used about 20% fewer tool calls and 25% fewer tokens than version 1.2. These numbers have not been confirmed by independent tests. Still, they measure something users can feel: waiting time, price, and the number of possible errors.
Long agent tasks rarely follow a perfect plan. A user may add new facts or change the goal. Meta says the model is better at connecting a new message to the right job. If this works, people will need to repeat less information.
What users should measure
Meta also says the model is safer against harmful instructions hidden in web pages. Outside experts still need to test this. Agent safety also depends on permissions, tool design, clear approval screens, and a good record of every action.
Users should count completed tasks after human review. They should also count retries, wrong actions, and the time needed to correct the agent. Meta says open model files may come later, but it has not given a date.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
We compared Meta’s launch article with independent reporting about access, price, future plans, and safety. The efficiency and safety numbers are Meta’s own results because no independent test was available when the article was published.