Gemini back in the day when version 2.5 could create stories up to 200 chapters or even 300. But now it can’t even last 20 chapters, the AI is already gasping for breath. They say the newest model, they say the best model, where is the proof?
Since creative writing is my niche, I’ve observed this overall phenomenon in models.
Task bias → pressure to finalize and complete the task → better benchmark results.
At the same time, “task bias” kills everything that requires a process, iterative building/creating, lack of rush, and enjoying the sheer act of building itself. It totally kills the possibility of writing a novel.
Same here. Its very sad to see that gemini 2.5pro deprecation.
Did you be able to solve this? Find something or some ways to reach to that level of performance? I would really appreciate that if you can share if there is any possible.
Ive been looking to everywhere and been trying other models too but havent find any model that performing the same level or havent find any solution that to get same level of creativity, reasoning, deep thinking, writing performance and intelligence etc. somehow with some any other ways. There are many people complaining and mentioning and trying to pull an attention to this problem but also no productive news or help from google or teams from gemini.
Where and in what?
I achieved the best results in gemini-cli (when I had full control over the context) and the model didn’t have strange filters. AiStudio before the filter era was also okay.
Currently I use my own app or claude-code. Claude-code offers the convenience of context management through the rules folder, allowing you to load context to the brim.
Model?
Gemini-3.0+ models are currently a waste of time for me when it comes to creative writing. If I use them at all, it is for translations or handling small tasks that are simply tedious. I will wait for 3.5 (though I have no hope, as there can’t be much difference in tasks that aren’t a priority for Google), maybe version 4.0 will change something? Although I have concerns, since the Gemini model ends up in the search engine and the Gemini app, which forces it to have a ‘task bias’, and that will kill creative writing.
Alternative?
It won’t be a surprise, but Fable impressed me as an amateur writer. GLM-5.2 was also doing quite well. Yesterday I tested Muse 1.2 and I am positively surprised.
For an intellectual depth and creativity the solution is fable 5, for an intellect and analytic accuracy the solution is opus 5.
So the best way to get the highest performance from them is using on claude code with cli versions. Also I believe finding a way to trick safety layers (traditional jailbreak is not the solution actually more hurting the logic and vocabulary in my experience), cli structural bypass can help to get some good performance. I wonder can anthropic developer console can work as aistudio and provide no filtered versions of the models.
Also will check the muse 1.2. Sounds promising.
Thank you for the answer btw.