admin The company said the model, released on Wednesday, is designed to handle longer workflows while making fewer unnecessary moves. Muse Spark 1.3 remains priced at $1.25 per million input tokens and $4.25 per million output tokens through Meta Model API, preserving the rates used for Spark 1.2 as Meta presses its effort to attract developers building AI agents and coding systems.
Meta said Spark 1.3 was trained on a set of long-horizon coding tasks and is intended to work more efficiently in common software-engineering workflows. In internal comparisons conducted by Meta engineers, the new model used about 20% fewer tool calls and roughly 25% fewer tokens than Spark 1.2, while also producing less verbose output and cleaner coding style.
The company also said Spark 1.3 is better at sustaining extended agentic work in which a model must collaborate with a user, manage workflows within one thread and gather information through tools. Meta said the system can build context from conflicting or incomplete material, identify gaps in its own plan and keep track of information acquired during a task before producing a final deliverable.
Those capabilities are central to Meta’s push into so-called agentic AI, where models move beyond answering prompts and instead take a sequence of actions using software tools, browsers or other services. The company has positioned Muse Code as a multi-agent terminal coding product built around the Muse Spark family, while Meta Model API gives developers direct access to the models for applications.
Muse Code was introduced alongside Spark 1.2 in August as a terminal-based coding agent capable of planning, implementing and validating changes across large software repositories. Meta has since expanded access to its developer tools as it competes with rival model providers for programmers, enterprise customers and developers building autonomous or semi-autonomous software agents.
Spark 1.3 also retains the multimodal capabilities of the Muse family, which can process material including images, video and documents alongside text. Meta says its visual reasoning operates through an execution environment rather than relying solely on fixed scripted steps, allowing the model to combine perception with tool use in workflows that may span multiple applications.
Meta said the new version includes safety improvements aimed at agentic and coding use. The company said Spark 1.3 shows greater resistance to adversarial inputs and prompt-injection attempts, a concern for agents accessing external information or taking actions through connected tools.
The model has also been tuned to judge irreversible actions carefully during complex tasks, according to Meta. The company said the changes are intended to improve discretion in long-running workflows, where errors can become more consequential if an agent is allowed to modify files, interact with online services or execute commands without sufficient checks.
Meta is initially making the reasoning modes already available in earlier Spark versions accessible with Spark 1.3. A maximum-reasoning mode is due later after additional safety testing, the company said, indicating that some higher-compute capabilities are being held back pending further evaluation.
The release comes as technology companies intensify competition around coding models and software agents, an area viewed as one of the clearest commercial uses for advanced generative AI. Developers increasingly compare models not only on benchmark scores, but also on cost, latency, tool reliability, context handling and the amount of human intervention needed to complete extended tasks.
Meta has sought to differentiate the Muse family partly through pricing and integration. Its developer API supports agentic workflows and structured outputs, while Muse Code is built to coordinate multiple agents inside a terminal environment. The company says the model can support tasks such as software development, computer use, search grounding and multi-agent orchestration.
