Muse Spark 1.3: Meta Is Back at the Top, and the Best Open-Weight Model Could Follow

Set Trending Topics as a preferred source on Google.

Mark Zuckerberg announced the rollout of Muse Spark 1.3 on X, and this time the numbers back him up. The model delivers “frontier performance that is almost too cheap to meter,” the Meta CEO writes, calling it “the biggest jump we have made yet on coding and agentic work.” It is available in Muse Code, Meta’s answer to Claude Code and OpenAI Codex, as well as through the company’s own API. The post ends with a line that will keep the industry busy for a while: up next are a model teased with a watermelon emoji and open-weight releases of Muse Spark.

Only Anthropic Still Ranks Higher

How far Meta has come is clear from the evaluation by Artificial Analysis. This is the fourth Muse Spark release in five months, and the trajectory is steep: the publicly available variant, Muse Spark 1.3 (xhigh), scores 61 on the Intelligence Index, up from 57 for version 1.2 and 53 for version 1.1. That puts it level with GPT-5.6 Sol (max), Grok 4.6 (high) and Claude Opus 5 (high).

The stronger “max” variant, currently in limited preview for Meta’s partners, reaches 62. Only two models rank above it in the overall index, both from Anthropic: Claude Fable 5.1 (max) at 66 and Claude Opus 5 (max) at 63. A year after the Llama 4 generation fell off the pace, Meta is back among the three strongest providers of AI models.

Just as notable is who now sits behind Meta. The highest-placed Google model in the index is Gemini 3.8 Flash (high) at 59 points, and no current Pro model from the Gemini line appears there at all. OpenAI and xAI, in their strongest listed configurations, do not clear Muse Spark 1.3 either.

The Gains Come From Agentic Work

According to Artificial Analysis, the jump over version 1.2 comes mainly from agentic tasks and scientific reasoning:

• Tau3-Bench Banking: from 35 to 47 percent (xhigh) and 52 percent (max). The max variant currently holds the top score of any model tested.
• Terminal-Bench 2.1: from 80 to 85 percent (xhigh) and 86 percent (max).
• GDPval-AA v2: Elo rating from 1,615 to 1,709 (xhigh) and 1,754 (max).
• Scientific reasoning: CritPt climbs from 18 to 26 percent, GPQA Diamond from 90 to 94 percent, with Humanity’s Last Exam and SciCode each adding two to three points.

Meta pays for the max variant’s stronger agentic results in compute, since it reasons 62 percent longer on GDPval-AA v2 and 28 percent longer on Tau3-Bench Banking than the xhigh version. There are two regressions as well: AA-LCR drops from 83 to 79 percent, and accuracy on AA-Omniscience falls by three points. Artificial Analysis attributes the latter to the model declining to answer more often when it is unsure, which also brings its hallucination rate down.

Then there is price. Muse Spark 1.3 (xhigh) costs $0.55 per Intelligence Index task, and no model scoring 59 or above comes in cheaper. Its direct peers land at $0.94 (Grok 4.6 high), $0.95 (GPT-5.6 Sol max) and $1.23 (Claude Opus 5 high). Meta’s rates are unchanged from version 1.2 at $1.25 per million input tokens and $4.25 per million output tokens, with cache hits at $0.15. The context window holds one million tokens, and the model takes text, image and video as input.

The Real Leverage Sits With the Open Weights

The situation gets interesting where Meta could play to its old strength again. Chinese labs currently lead the open-weight rankings: Kimi K3 (max) from Moonshot AI and GLM-5.3 (max) from Z.ai both sit at 60 points, with Alibaba’s Qwen3.8 at 58. A model at the level of Muse Spark 1.3 would take that lead outright and would be the first open model to stand level with the closed systems from Anthropic and OpenAI.

That is also where Zuckerberg’s announcement leaves room for interpretation. What Meta has committed to so far is releasing the weights of Muse Spark 1.2, and at 57 points that version would land behind Kimi K3, GLM-5.3 and Qwen3.8. Whether the weights of version 1.3 will follow is still undecided. That single decision determines whether Meta’s comeback also reshapes the open-source field.