Meta closes the gap
Meta released Muse Spark 1.3 on September 2, the fourth update to its Spark model family in five months and, the company says, its largest single jump in performance so far. The model is live immediately in Muse Code, Meta’s terminal-based coding agent, and through the Meta Model API, according to Meta’s official announcement.
Smaller, faster, closer to the frontier
Independent testing from Artificial Analysis put the model’s “xhigh” configuration at 61 points on its Intelligence Index, up from 57 for the prior Spark release, with a “max” preview variant reaching 62. That puts Muse Spark 1.3 roughly level with OpenAI’s GPT-5.6 Sol (max) and Anthropic’s Claude Opus 5 (high) on the same scale, though still behind Anthropic’s top-tier Claude Fable 5.1 (max), which scores 66.
Meta says the bigger gains are in efficiency rather than raw benchmark rank: the model uses about 20% fewer tool calls and 25% fewer tokens than its predecessor on comparable coding tasks, which Meta attributes to training on more long-horizon coding sessions. Pricing for the xhigh tier is unchanged at $1.25 per million input tokens and $4.25 per million output tokens, with cached tokens discounted to $0.15 per million — a cost-per-task figure Artificial Analysis calculated at roughly $0.55, below GPT-5.6 Sol (max) and Grok 4.6 (high).
Built for longer-running agents
Meta’s chief AI officer, Alexandr Wang, described the release as the company’s biggest jump in model performance to date, framing it as another step toward AI agents that can work on a user’s behalf for extended stretches without supervision. According to Meta, the model now asks clarifying questions on ambiguous prompts, flags irreversible actions before taking them, and holds context across multiple linked tasks in a single thread — capabilities aimed at the kind of long-running, semi-autonomous coding sessions Meta has been building toward since launching Muse Code.
Meta also said it strengthened the model’s resistance to prompt injection and adversarial inputs as part of the release, without giving further technical detail. A larger model Meta has been training, code-named Watermelon, remains in development and unreleased. Meta’s Muse Spark 1.1 launch in the spring marked its first major push into AI coding tools.
Developers can access Muse Spark 1.3 by signing up at dev.meta.ai.