The Pulse
Meta cuts Muse Spark coding tool calls by 20% in version 1.3
Meta released Muse Spark 1.3 on September 2, 2026, with improvements for long-running agent workflows and software engineering. The company says internal comparisons show about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.
AI.info Team ·
Meta says Muse Spark 1.3 completes coding work with about 20% fewer tool calls and 25% fewer tokens than its predecessor, according to internal comparisons by Meta engineers. The September 2 release targets the part of software development where AI systems often lose track: long tasks that require planning, repeated tool use, changing instructions and checks against a final goal.
Muse Spark 1.3 is available through Meta’s Muse Code terminal agent and the Meta Model API. Meta presents the update as a practical refinement of Muse Spark 1.2 rather than a new product, with less unnecessary back-and-forth, shorter responses and what the company describes as cleaner coding behavior.
The release arrives as Meta builds toward personal agents that can work across applications and maintain context over extended sessions. Alexandr Wang, Meta’s AI chief, told Axios, “It’s very competitive with frontier models,” while describing the model’s usability improvements as relevant to personal agents that could work on a user’s behalf.
Wang’s comments to Axios place the release in a wider push by Meta to narrow the gap with Anthropic, OpenAI and Google in coding and agentic software.
20% fewer tool calls on engineering work
Meta says Muse Spark 1.3 was trained on more long-horizon coding tasks and performs better in common engineering workflows. Relative to Muse Spark 1.2, the company reports that the model uses about one-fifth fewer tool calls and one-quarter fewer tokens in internal comparisons. Those figures are not an independent benchmark, but they point to the operational trade-off Meta is emphasizing: an agent that reaches a result with less interaction can reduce latency and usage costs.
The model is designed to plan across larger tasks, inspect messy sources, identify gaps in its own work and preserve information needed for a final deliverable. Meta says it can juggle several workflows in one long thread, including situations in which a user returns to an earlier request, interrupts an active task or changes direction.
That focus builds on Muse Code, which Meta introduced with Muse Spark 1.2 on August 5. Muse Code was designed to work across large repositories, plan changes, write code and validate results. Its runtime records model calls, tool runs, approvals and edits in a local event log so a task can resume after a failure.
Meta trains the model to stop and ask
Meta’s release puts user interaction alongside raw task completion. Muse Spark 1.3 is trained to ask clarifying questions when instructions are ambiguous, request help when it gets stuck and seek confirmation before taking consequential actions. For long tasks, Meta says users can choose between frequent progress updates and quieter background work.
The model also aims to make fewer unsupported claims about its own performance. Meta says it has improved the system’s awareness of what it knows, what it can do and when it has hit a barrier. In an agent that edits files, runs commands or acts across multiple services, recognizing a limitation can matter as much as producing a correct answer.
Meta separately says the new model has stronger resistance to adversarial inputs and prompt injections. It also claims better calibration around irreversible actions. The company does not publish a detailed incident rate or an independent safety evaluation in the release post, so those statements remain company-reported performance claims.
The evaluation compares long tasks, not just code snippets
Meta’s accompanying evaluation methodology covers professional work, computer use, web research, automation, software engineering, code understanding, long-context retrieval and instruction following. The document says Meta compared Muse Spark 1.3 with Muse Spark 1.2, Claude Opus 5 and GPT-5.6 Sol where comparable results were available, using maximum reasoning effort for the newer model and the two external systems.
The coding tests include DeepSWE v1.1, a 113-task benchmark covering 91 repositories and five programming languages; SWE-Atlas Codebase QnA, which tests comprehension across 124 tasks and 11 production repositories; and Terminal-Bench 2.1, which contains 89 terminal tasks. Meta says coding environments do not have external internet access unless a benchmark specifically permits it.
Those comparisons require care. Meta’s methodology notes that third-party model runs use common settings where practical, but prompts, tools and runtimes may not reflect the best provider-specific configuration. The company also reports the highest comparable primary-metric result available from its own evaluation, an official leaderboard or a model provider’s published result.
Muse Code and Meta Model API get the release
Muse Spark 1.3 is available in Muse Code and through the Meta Model API, according to Meta’s release. Developers can install Muse Code on macOS or Linux with Meta’s published command:
curl -fsSL https://dev.meta.ai/install.sh | bash
The immediate audience is developers running agents against repositories, terminal environments and structured business tasks. Meta’s examples also show the model handling mixed professional work, including engineering reports, audio editing instructions, presentations and constituent-feedback summaries.
Meta says its roadmap includes larger models and an open-weights release for Muse Spark. The September 2 announcement does not provide a date for either. For now, the concrete change is narrower: Muse Spark 1.3 is Meta’s latest available model for Muse Code and its API, with the company claiming fewer tool calls, fewer tokens and better control over long-running work.