Gemini 3.6 Flash: Don't Believe the Benchmarks
You will know what Google's latest Gemini Flash release changed, how to check the benchmarks yourself on Artificial Analysis, and when I would still use Gemini instead of Grok or GLM.
Was Sie brauchen
A web browser
Optional: a Gemini, Grok or GLM (Z.ai) account to test the models yourself
Schritt für Schritt
Know what was released
Google released three models at once: Gemini 3.6 Flash, 3.5 Flash Lite and 3.5 Flash Cyber. Flash Cyber is not really a new model, just a version of Flash aimed at cybersecurity. Expect every release to claim it is better and faster than the last generation.
Understand the Gemini lineup
Gemini comes in three sizes. Flash Lite is the smallest and cheapest, good for analysis at scale. Its main strength: Gemini can still analyze videos, documents and images natively, with very strong OCR. Pro is the largest, but no 3.5 Pro or 3.6 Pro has shipped since 3.1 Pro.
Compare 3.5 Flash and 3.6 Flash yourself
Open Artificial Analysis, the source of the charts in the video. Compare 3.5 Flash and 3.6 Flash on intelligence: they sit at the same level, so the new version brings no real performance gain. 3.5 Flash Lite ranks far lower, which is normal for a smaller model.
Artificial Analysishttps://artificialanalysis.ai
Check Gemini against the top models
On the same site, put 3.6 Flash next to the current top models (Kimi K3, GLM, Fable, Sol, Grok). 3.6 Flash scores 50 on intelligence, while the top five score 10 to 20 percent higher. 3.6 Flash is the fastest, but for building apps you want the best quality at around 60 tokens per second, not maximum speed.
Look at cost per task, not price per token
Open the cost per task view: it shows what it actually costs to get work done. Grok 4.5 costs about 40 percent less than Gemini 3.6 Flash and scores higher on intelligence, close to Opus 4.8. That makes it both better and cheaper for this comparison.
Pick what to use instead
Skip Gemini 3.6 Flash. Use Grok for better quality at lower cost, or GLM if you want an open source model you can run yourself. Keep Flash Lite only for cheap, large scale analysis of documents and videos, knowing nothing has really changed with this release.
Grokhttps://grok.com
GLM (Z.ai)https://z.ai
Worauf Sie achten sollten
Vendor benchmarks always show the new model beating the last one. Check an independent source like Artificial Analysis before switching.
Speed alone is a poor reason to choose a model: for applications, quality at around 60 tokens per second matters more.
Judge cost by cost per task, not by headline pricing. A cheaper-looking model can cost more to finish the job.
Do not wait on Gemini Pro: no 3.5 Pro or 3.6 Pro has been released, and there is no sign of when it will be.
