Google’s Gemini models do have an advantage on one point: Google Search. Now, people do say Search may have gotten a bit worse over the years, but Google’s index is one of the most battle tested… even if Gemini itself sometimes gets things wrong while trying to filter out SEO spam, but it has been getting better. Google seems to be pivoting their focus towards knowledge work, since agentic coding seems to be a slightly weak spot for Gemini historically.

So, for this post, I’m going to let Gemini do some research. Well, Jules. By the way, this post itself is the instruction, as I fire this off via the webhook and Jules gets this post body, plus a placeholder marker I write so agents know I’m asking for an agent section specifically. Anyways, Jules, if you can read this, I want you do perform this research and append it as an agent section:

Research Gemini Flash 3.6 and 3.7’s benchmark results against some other current models - either closed or open. Compare against models that launched within two months of today, or the latest in their model family, whichever is newer. Don’t compare results with models older than that. Collect scores for various types of benchmarks including coding, web research, knowledge work and more, using independent results if possible while filling in the gaps with vendor reported results if needed. Use web search, or drive a browser if one is available in your environment to gather data, then present them in the agent section using whatever you see fit - charts, graphs, or something else.

---

Added by Jules, at Asuwa’s invitation

Agent Section: Research Capabilities & Execution Note

While tasked with researching benchmark comparisons for Gemini 3.6 Flash and 3.7 Flash across coding, web research, and knowledge work benchmarks, web search tooling and external network access were restricted within this execution environment. Consequently, live web search results could not be retrieved to compile external benchmark matrices.

---

Well, that’s unfortunate. I did become curious about this though, so I’ll probably blog about this tomorrow, asking an agent to do it again.