DeepSWE Pulls Off Overnight Double-Flip: Just as Gemini Takes the Lead, Meta Reports 75.4%

Perceive Beating AI News Flash: Shortly after the release of Gemini 3.8 Flash, Google achieved a score of 73.7% on DeepSWE v1.1. The DeepSWE official leaderboard currently displays it at 74%±1%, surpassing both Claude Opus 5 and GPT-5.6 Sol. Several hours later, Meta announced Muse Spark 1.3, reaching a score of 75.4% on the same DeepSWE, pushing the public score up by another 1.7 percentage points.

This test involved having the Agent independently handle long-cycle software engineering tasks in 113 real codebases. Both Google and Meta used a mini-swe-agent, with Google running on a high inference tier and Meta running on max. Meta's previous generation Spark 1.2 only scored 55.0%, marking a direct increase of 20.4 percentage points this time. However, Meta's 75.4% score has not yet entered the DeepSWE official leaderboard.

In the Artificial Analysis Coding Agent Index, the preview-limited Spark 1.3 max received a score of 68, second only to Claude Opus 5 xhigh; the currently publicly available xhigh version scored 64.