>from the looks of it, 3.5 Flash is still better than most models

Define "better". I guess it depends on what you're using it for. I use it almost daily as an alternative to google search and it's great for that, but I think it's absolute garbage for coding and reasoning.

For questions related to coding, solving Arch Linux and WINE Lutris issues, helping me with MXLinux issues, and wifi issues on an old rooted huawei tablet running LineageOS, it was consistently wrong, constantly giving out confident but outdated or misinformation, or hallucinating stuff while gaslighting me. Every time I would point out it was wrong, it would re-check and keep apologizing and then repeat giving me wrong answers, and then apologising again and so on. It doesn't matter what prompts or jailbreaks you give it to get 3.5 Flash to chew longer on complex problems for better reasoning and accuracy, it just defaults to being lazy and giving you the quick and easy answer from its weights, which can be totally wrong. Same for asking it to write me a cover letter based on my resume and the job description I wanted to apply to. It massively sucked at that too and made up a bunch of unusable fake sounding BS.

Basic free tier ChatGPT 5.5 would blow it out of the water on all of those tasks. Hell, even Grok free is better at that, it gave me a one-shot Arduino code that blew Gemini 3.5 flash away.

3.5 Flash seems tuned to just eyeballing basic answers to general purpose questions that resemble Google searches like "give me a recipe" or "give me a workout plan", or "what's the difference between Arch and Fedora based distros", not to solving complex issues that require cognition and accuracy. That's what the 3.1 Pro is better for according to Gemini. Oh and it is also gaslights you by starting the answers with first telling you how amazing things from your question are, which is insanely annoying but I guess Google's A/B testing found out the majority of Average Joe midwits love it when "the AI" reinforces their choices and decisions like a fake friend.

I think Google just doesn't care about being the SOTA for coding, reasoning and accuracy, since they're in the ads and search business for everyone, not in the agentic coding business for pro-sumers, so if the answers are some hallucinations that sound "good enough" to its clueless search user base, but is at least dirt cheap to run on their datacenter hardware, then it's already more than enough for them and they can all it a day.

Meanwhile OpenAI and Anthropic don't have search and ads monopolies, so they need to perform well at certain task for people and businesses to give them their hard earned money for them to survive. For them, nailing stuff like coding and writing accuracy is an existential threat, not a hobby sideproject like it is for Google.

> They're still crunching on it. No ETA given.

Thank God. I'd rather companies ship something when engineers say it's actually ready rather than when the suits want something to show on stage to pump their egos and career exposure but turn out to be a massive disappointment covered in fluff.

Although it does feel very embarrassing for Google who invented transformers and has more money than both Anthropic and OpenAI combined, to fall behind them at the LLM race.