Meta released Muse Spark 1.3 on 2026-09-02, framing it as improved on agentic and coding tasks and, in Meta’s words, easier to use in real-world settings after months of broad adoption of Muse Code and the Meta Model API. Meta describes the release as advancing its work toward personal superintelligence. The model with max reasoning is available on Muse Code and the Meta Model API through dev.meta.ai.
The efficiency claim is the concrete one: roughly 20 percent fewer tool calls and roughly 25 percent fewer tokens than Muse Spark 1.2, with a less verbose and cleaner coding style. For anyone paying per token on long agentic runs, that is a cost reduction independent of any change in the posted price, which Meta keeps at the Muse Spark 1.2 level.
Most of the announcement is about collaboration behavior rather than benchmarks. Muse Spark 1.3 is trained to ask clarifying questions on ambiguous prompts, invoke help from the user when it stalls, and confirm before taking consequential actions. On long-running work it adapts to preference, either giving frequent status updates or executing silently in the background. Meta also claims better multitasking within a single messy thread - mapping incoming prompts to the correct task whether the user is steering a past request or interrupting it - and better calibration on its own limits, so it flags hurdles instead of hallucinating an outcome. Meta reports stronger adversarial robustness, including improved resistance to prompt injection.
Meta’s published scorecard compares Muse Spark 1.3 against Muse Spark 1.2, GPT-5.6 Sol at max and Claude Opus 5 at max across agent, coding, instruction-following and long-context evaluations, with detail deferred to a separate report. The competitive framing is explicit and was echoed by Meta’s chief AI officer in press coverage the same day: this is Meta arguing it has closed most of the gap to the leading closed labs on agentic work.