Photo: Tima Miroshnichenko / Pexels
The latest update from Meta Superintelligence Labs cuts tool calls by 20% and token usage by 25%, sharpening its AI coding agent for complex engineering workflows.
Meta is shipping AI model updates the way most companies ship Slack messages: constantly, and with increasing urgency. On September 2, CEO Mark Zuckerberg announced the release of Muse Spark 1.3, the third major update to a model family that didn’t even exist five months ago.
The new version brings meaningful efficiency gains for developers, with a 20% reduction in tool calls and 25% fewer tokens consumed compared to its predecessor, Muse Spark 1.2. In practical terms, that means the model does more with less, producing cleaner code while burning through fewer computational resources to get there.
What Muse Spark 1.3 actually does differently #
The core upgrades center on two areas: agentic workflows and coding performance. Muse Spark 1.3 has been trained on a broader set of long-horizon coding tasks, which is the AI equivalent of teaching someone to run a marathon instead of just sprinting 100 meters.
The 20% reduction in tool calls matters because every tool call introduces latency and potential error. Fewer calls with the same output quality means the model has gotten smarter about when to reach for external tools and when to handle things internally.
Similarly, the 25% decrease in token usage translates directly to cost savings. Tokens are the billing unit for most AI APIs. A model that achieves the same results with a quarter fewer tokens is, all else equal, 25% cheaper to run at scale.
Muse Spark 1.3 is available immediately through Muse Code, Meta’s beta coding agent, and the Meta Model API. A “max reasoning” mode is also in the pipeline but won’t ship until it clears additional safety testing.
A five-month sprint from zero to version 1.3 #
The pace of iteration here is worth noting. The Muse Spark family debuted in April 2026. Version 1.1 followed in July. Version 1.2 landed in August. Now 1.3 arrives in September.
That’s four releases in roughly five months, a cadence that would be aggressive even by the standards of the current AI arms race.
The entire effort traces back to an uncomfortable realization inside Meta: its open-source Llama models, once a point of pride, weren’t keeping up with the frontier models from OpenAI and other competitors. Rather than incrementally improving Llama, Zuckerberg opted for something more drastic.
Meta Superintelligence Labs, or MSL, was created as a dedicated unit to rebuild the company’s AI capabilities from the ground up. The lab is led by Alexandr Wang, a notable choice given Meta’s $14.3 billion investment in Scale AI, the data labeling company Wang founded.
The Muse Spark models represent a philosophical shift for Meta’s AI strategy. Rather than competing purely on benchmark scores for general-purpose language models, Meta is betting on multimodal reasoning and tool utilization. The stated goal is to deliver a “personal superintelligence” experience, one where the AI doesn’t just answer questions but actively collaborates on complex, multi-step objectives.
The competitive landscape gets tighter #
For enterprise customers evaluating AI coding tools, a 20% reduction in tool calls and 25% fewer tokens aren’t abstract improvements. They show up directly in API bills, latency measurements, and developer satisfaction scores. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our