Meta announced that it released the AI model "Muse Spark 1.3" on September 2, 2026 (local time).
Muse Spark 1.3 is trained to handle long-term coding tasks. Meta stated that the model recorded scores exceeding GPT-5.6 Sol and Claude Opus 5 in benchmark tests measuring coding performance.
In tests conducted by the third-party organization Artificial Analysis, Muse Spark 1.3(max) surpassed GPT-5.6 Sol in the comprehensive intelligence performance index, matching the results of Claude Fable 5.
On the Agentic Index, a metric for agent performance, Muse Spark 1.3(max) recorded scores equivalent to Claude Opus 5, GLM-5.3, and Grok 4.6.
According to Meta, the model was developed with user collaboration in mind. It possesses the ability to ask clarifying questions when receiving ambiguous instructions or request additional guidance from the user when it encounters a bottleneck in processing.
Meta stated that compared to the previous generation, Muse Spark 1.2, redundancy has been reduced, allowing for the output of more refined code. Specifically, the company explained that tool calls decreased by approximately 20% and token counts by approximately 25%, improving processing speed and efficiency.
Muse Spark 1.3 is available via API. inference levels up to xhigh are already provided, and the max mode is scheduled to be available after the completion of safety testing.
API pricing is $1.25 per million input tokens, $0.15 for cached input, and $4.25 for output. Additionally, a "Contributor tier" is available, offering lower rates in exchange for providing prompts for training.
Meta CEO Mark Zuckerberg said that the model represents a significant leap forward in coding and agent tasks. He also explicitly stated that the Muse Spark series will be made available as open models in the near future.
Source: Metaが「Muse Spark 1.3」を発表、ついにAnthropicやOpenAIの最先端モデルに追いつく性能を達成 (GIGAZINE, 2026-09-03)