Google's Gemini 3.5 Flash is based on the Gemini 3 Flash architecture according to its system card, explaining a price increase driven by higher token speeds and interactivity rather than added scale.

Google is still behind. Previously I thought Gemini 3.5 Flash was more expensive because they trained larger models for the 3.5 series, but Google said in their Gemini 3.5 Flash system card that it's based on Gemini 3 Flash, meaning that the price increase is likely because of the high interactivity (tok/s/user). That also implies that Gemini 3.5 Pro will not be a larger Mythos sized model. Gemini 3.5 Flash is currently between GPT-5.4 and GPT-5.3 and between Opus 4.6 and Opus 4.7 on the Artificial Analysis Index, implying a ~2-3 month lag, but this includes a bunch of saturated/bad benchmarks. On WeirdML it's lagging behind GPT-5.2 and Opus 4.5. On SWE-Bench-Pro it's behind GPT-5.2 and on TerminalBench Hard it's worse than Opus 4.5. That is indicative of a ~5-6 month lag. Of course this is only the Flash model, but the Pro model won't be here for another month and you also have to consider that Anthropic had Mythos avaialble for internal use since February 24. Gemini 3.5 Pro will likely score above 60 on the Artificial Analysis Index, but will still be behind ~2-4 months on more important bechmarks and in real world coding use cases. Once we have the next version of GPT and Opus and Gemini 3.5 Pro we will know whether Google is falling further behind or not. From personal use I can say Gemini 3.5 Flash still feels like a smaller model as it misses the intent of my questions much more often than Opus 4.7 and GPT-5.5. The $9 output pricing also hurts the models position, as it didn't improve on reasoning efficiency as GPT-5.5 did over GPT-5.4