AI
Meta's Muse Spark 1.3 overtakes Google's latest Gemini in coding benchmarks hours after launch
Meta released its latest reasoning model with gains in coding performance, briefly unseating Google's new low-cost Gemini 3.8 Flash from the efficiency frontier.
The New Stack
Meta has rolled out Muse Spark 1.3, an updated version of its low-cost reasoning model, claiming substantial improvements in coding and agentic AI tasks. The release arrived just hours after Google published its own budget-friendly model, Gemini 3.8 Flash, creating a brief but notable moment in the competitive AI landscape.
Performance gains and benchmarks#
The new Spark iteration demonstrates considerable progress compared to its predecessor. Meta CEO Mark Zuckerberg described the update as representing "the biggest jump we've made so far on coding and agentic work," with benchmark results showing a 75.4% score on the DeepSWE coding benchmark and 98.1% on long-context evaluation tasks.
However, these comparisons come with important context. Meta assembled the benchmark table itself, drawing competitive data from a mix of its own assessments, public leaderboards, and figures reported by rival companies. The company also ran Spark 1.3 in a new "max" reasoning mode for headline comparisons, though this remains in limited testing while Meta completes safety reviews.
Independent testing from Artificial Analysis provides a clearer outside perspective. Their analysis shows Spark 1.3 at the standard "xhigh" reasoning level scores 61 on their Intelligence Index—matching GPT-5.6 Sol and Grok 4.6 at that capability tier. Crucially, Spark 1.3 achieves this performance at approximately $0.55 per task, making it the most cost-efficient model in its capability range. The max reasoning version scored 62, placing it among the top performers but still behind Claude Opus 5.
The brief Google-Meta showdown#
Google's Gemini 3.8 Flash launched mere hours before Muse Spark 1.3, initially claiming a spot on Artificial Analysis's Pareto frontier—the group of models with no superior alternative offering both better capability and lower cost. Google priced the model at $0.75 per million input tokens and $3.75 per million output tokens.
The competitive moment proved fleeting. Muse Spark 1.3 xhigh surpassed Gemini 3.8 Flash on both intelligence and cost metrics almost immediately, pushing Google's model off that efficiency frontier. Observers noted the remarkably short window, with one researcher remarking that "Google was ahead only a few hours," and another describing the turnover as lasting just 3.5 hours.
