Meta Muse Spark 1.3 Released: 25% Efficiency Gain, Tops Benchmark Charts

Meta has launched Muse Spark 1.3, its next-generation flagship AI model focused on code generation and long-horizon agent tasks — delivering significant efficiency improvements with no price increase.

Muse Spark 1.3 Performance Improvements: 20% Fewer Tool Calls, 25% Lower Token Usage

Compared to version 1.2, Muse Spark 1.3 achieves two key improvements while keeping API pricing unchanged at $1.25/$4.25. According to Meta's internal benchmarks, the model reduces tool calls by approximately 20% on coding tasks while cutting token consumption by roughly 25%. Developers get more execution efficiency at the same cost, effectively lowering the real-world usage expense.

This kind of optimization is especially valuable in development scenarios that require frequent model calls. Fewer tool calls mean faster response times, fewer intermediate steps in complex tasks, and greater overall workflow stability.

Coding Benchmark Results: DeepSWE 75.4, Terminal-Bench 88.8

Muse Spark 1.3 has posted strong results on widely recognized coding benchmarks:

• DeepSWE v1.1: Scored 75.4, surpassing the previous top result on the leaderboard
• Terminal-Bench 2.1: Reached 88.8, on par with leading models

These two benchmarks evaluate a model's capabilities in software engineering tasks and terminal command execution respectively — both are key indicators of an AI coding assistant's practical utility.

Notably, these scores were achieved after the model was specifically optimized for agentic work and long-horizon task planning. Compared to traditional single-turn coding assistants, Muse Spark 1.3 is better suited for complex development tasks that require multi-step reasoning and sustained context understanding.

DeepSWE is a benchmark designed to evaluate AI models in real-world software engineering scenarios, covering tasks like bug fixing, feature implementation, and debugging within complete code repositories — making it far more representative of production environments than simple code completion tests. Terminal-Bench focuses on a model's ability to execute complex instruction sequences in command-line environments, including file operations, environment configuration, and script execution, testing the model's understanding and control of OS-level interactions. Both benchmarks emphasize multi-step, stateful task execution rather than isolated single-turn Q&A, making them particularly well-suited for evaluating long-horizon agent models.

How to Use Muse Spark 1.3: Two Access Options

Muse Spark 1.3 is now officially available through two channels:

• Muse Code: Meta's integrated development environment, offering a seamless upgrade experience for existing users
• Meta Model API: Open to third-party developers for flexible integration into existing toolchains

Meta also revealed that the model's "Max reasoning" mode will launch after safety testing is complete. This mode implies deeper logical reasoning capabilities, potentially providing stronger support for complex algorithm problems or architectural design tasks.

The "Max reasoning" mode corresponds to the "extended thinking" paradigm that has gained traction in the large model space — where the model engages in a longer internal reasoning process before generating a final answer, typically expressed as more intermediate reasoning steps or self-verification loops. OpenAI's o-series and Anthropic's Claude 3.7 Sonnet have both introduced similar features. This mode tends to perform best on mathematical proofs, complex algorithm design, or system architecture decisions, though it comes with higher token consumption and longer response latency, requiring a trade-off between task complexity and cost.

Muse Spark 1.3 Pricing and Market Competitiveness

As Meta's flagship product in the AI coding space, Muse Spark 1.3 takes a competitive pricing approach. The $1.25 input price and $4.25 output price sit in the mid-range of mainstream commercial models, but when combined with improvements in token efficiency and call overhead, the effective cost of use becomes even more attractive.

The AI coding assistant market is intensely competitive. Products like GitHub Copilot and Cursor have built solid user bases, while Anthropic's Claude and OpenAI's GPT series continue to strengthen their coding capabilities. Muse Spark 1.3's differentiator lies in its dedicated optimization for long-horizon agent tasks — a unique advantage in development scenarios that require sustained, cross-session work.

From a product strategy standpoint, Meta is maintaining both the Muse Code environment and API services simultaneously — building an independent developer ecosystem while providing enterprise-grade integration capabilities. Whether this dual-track approach can carve out meaningful space in an already mature market will depend on real-world adoption by the developer community.