Meta says Muse Spark 1.3 closes the gap with Claude and GPT

Meta released Muse Spark 1.3 on September 2, 2026, the latest version of its flagship model line, available immediately through the Meta Model API and Muse Code, Meta's terminal coding agent, on macOS and Linux. Per Meta's announcement, the model is built for longer-horizon agentic work: sustaining multiple workflows in a single thread, asking for clarification on ambiguous prompts, and confirming before it takes consequential actions.

Efficiency is the headline

Meta's benchmark claim is specific and testable: roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on the same agent, coding, instruction-following, and long-context evaluations, benchmarked against 1.2 alongside GPT-5.6 Sol and Opus 5 (max). Meta lists safety changes as part of the same release: stronger adversarial robustness, better resistance to prompt injection, and improved calibration on irreversible actions.

SiliconANGLE's Mike Wheatley reports that Artificial Analysis, an independent evaluator, scored the model 62 on its Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5 and ahead of everything else. Meta Chief AI Officer Alexandr Wang told Bloomberg the model is "competitive" with Fable 5.1 and "better than" GPT-5.6 Sol at code generation, and said developers using the Muse Spark family are already burning through "trillions of tokens per week."

Pricing is unchanged from 1.2. Meta has not said whether it will release the weights, a notable silence given its earlier open-source framing for the Muse and Llama lines. SiliconANGLE flags this explicitly.

The efficiency number matters more than the leaderboard position. A model that does the same job with fewer tokens and tool calls costs less to run at the same price, which is the cost signal Merit AC's spend tracking is built to separate from capability marketing.