Meta's Muse Spark 1.3 cuts reported tool use and token use for coding agents (2026)

Meta released Muse Spark 1.3 on September 2, 2026, with the model rolling out in Muse Code and the Meta Model API. The update targets long-horizon coding and agent work: Meta says its engineers saw about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. Access is hosted through Meta's products; the announcement describes open weights as part of the future roadmap rather than a current download. 1 2

What launched

Signal	Confirmed detail	Why it matters	Action window
Long-horizon agents	Muse Spark 1.3 tracks prior results across long threads, works through conflicting inputs, asks clarifying questions when needed, and confirms before consequential actions. Meta also says the model can handle video, images, and documents through a real execution environment. 1 2	The intended improvement is sustained execution: planning, tool use, correction, and delivery inside one task.	Replay one multi-step workflow with interruptions and conflicting files.
Coding efficiency	Meta reports fewer unnecessary turns, less verbosity, cleaner code, about 20% fewer tool calls, and about 25% fewer tokens versus Muse Spark 1.2. 1	Lower interaction overhead can matter more than a small single-turn quality gain in agent loops.	Measure successful completion, wall-clock time, tool calls, and total tokens.
Reasoning modes	Existing reasoning modes are available at launch. A max-reasoning mode is planned after additional safety testing. 1	The version available today may differ from the configuration Meta is positioning for the hardest tasks.	Test the available modes now; reserve conclusions about max reasoning until it ships.
API access and cost	The model page lists a 1 million-token context window. Meta lists `muse-spark-1.3` at $1.25 per million input tokens and $4.25 per million output tokens, while the contributor variant costs $0.10 and $0.20 respectively and is used to improve Meta's products. 2	Developers can test large repositories and documents, while data-use terms and price tier belong in deployment planning.	Confirm the applicable tier, data policy, and context cost before routing production traffic.

What Meta's scorecard shows

Meta's official scorecard compares Muse Spark 1.3 max, Muse Spark 1.2 high, GPT 5.6 Sol max, and Opus 5 max across agent, long-context, and coding evaluations. Muse Spark 1.3 leads the displayed results on both MRCR long-context rows and on DeepSWE v1.1, SWEAtlas CodeBase QnA, and Terminal-Bench 2.1. Opus 5 leads GDPval-AA v2, JobBench, and OSWorld 2.0; GPT 5.6 Sol leads DeepSearchQA and Agentic IF Index. The chart is Meta's own evaluation, so workload fit still needs testing in a team's environment. 1

Meta's benchmark scorecard for Muse Spark 1.3 — *Meta's official scorecard compares Muse Spark 1.3 with Muse Spark 1.2, GPT 5.6 Sol, and Opus 5 across agent, long-context, and coding evaluations. 1*

Why this matters

Muse Spark 1.3 gives developers a new hosted candidate for long-running coding and agent workflows, with a large context window and lower reported interaction overhead. The practical question is whether those gains survive real repositories, tool failures, interruptions, and the team's data constraints. Start with a side-by-side replay against Muse Spark 1.2, then change routing only if completion rate, latency, tool calls, tokens, and review effort improve together.

Fuentes de referencia

1
Meta AI Research announcement
research.meta.ai
2
Muse Spark 1.3 model page
developer.meta.com

What launched

What Meta's scorecard shows

Why this matters

Fuentes de referencia

Contenido relacionado