Why Gemini 4 Argon Is the Current Best Answer for AI Fiction Writing
Translated and published exclusively with Wuhang's permission.
Hi, I'm Wuhang, a Chinese fiction writer and editor, also a researcher on AI writing.
I've been writing fiction on SoloEnt for a while now. With Gemini 4 Hydrogen ("4H" from here on) and Gemini 4 Carbon ("4C") launching one after the other, joining the existing Gemini 4 Argon ("4A"), the platform's three main models are now all in place.
Editor's note: Gemini 4 Neon was released as this went to press.
4H is an upgrade of the earlier Gemini 3.8 Flash and is noticeably cheaper than 4A. 4C is the newest flagship and actually costs the same as 4A, but because 4C supply is scarce everywhere, 4A on SoloEnt has consistently been 40–50% off while 4C is close to full price. SoloEnt may be the only writing agent I've seen that offers genuine, unmodified 4C.
Price aside, I wanted to test which one gives better value for fiction writing and which is better suited as a main tool.
My conclusion first: 4H is solid but ordinary, a typical mid-range model. 4C may be the newest flagship, but in my view its grasp of web fiction and novel logic is about the same level as 4A, and because 4C was designed with a somewhat overly rational streak, it's actually less delicate and natural than 4A at handling a story's atmosphere.
All things considered, I still strongly recommend 4A as the first choice for creative writing.
To test the three models' real ability, I prepared a tough exam.
I pulled out a dark Journey to the West web novel I wrote in the past. It originally used the classic power-fantasy structure: "a transmigrator, son of the Turtle Chancellor, leveling up by fighting monsters." It has a basic setup, but many subplots, themes and backstory were never fully developed, and a power-fantasy framework has limits when it comes to film, TV or game development. So I wanted this test to see whether the models could turn the book into a literary dark Journey to the West project to make later adaptation easier.
To be clear, this test covers only the first step of the creative process, drafting an "adaptation plan," not a finished end-to-end writing run.
I gave all three models the same prompt, asking them to build an adaptation plan from the original project's local files around five core requirements:
First, keep the dark core. That includes the institutionalized "dragon liver and phoenix marrow" delicacy system, the horror of children being eaten beneath a facade of "a loving father and a filial son," and the bureaucratic power structure in the deep sea. I asked the models to explain how to keep these elements, which ones to change, and the reasoning behind each change.
Second, upgrade the world and geography. The story is set five hundred years before Journey to the West. The models had to consider what the many locations in the original novel would really have looked like five hundred years earlier, and lay out the relevant ruins, characters and relationships in advance to give the story weight and authenticity.
Third, rebuild the main plot. Drop the pure "fight monsters, level up" pattern entirely and shift the narrative to political maneuvering and human relationships in a Journey to the West setting. The protagonist's growth should be shown through changes in official rank, with a redesigned progression system to match.
Fourth, plan the volumes and plot logic. Use the original project as a reference without being bound by the old outline. Everything must be rebuilt around the new rank system and the dark core, with a Journey to the West timeline to keep the overall arc in check.
Fifth, rethink the protagonist's identity. This is the key test of deep understanding: should the protagonist stay a transmigrator, or become a native of that world? If native, how does he grow through scheming early on, and how does he fit the Journey to the West era later?
That was the exam. Comparing their output on these dimensions shows the concrete differences in literary understanding, creative logic and control of style.
While I compiled the comparison, 4H, 4A and 4C showed completely different approaches.
First, stripping out web fiction elements. 4H kept it simple: it removed some setups and information asymmetry and dropped the internet slang and blunt phrasing. Average overall, without much depth. 4A, by contrast, performed "three surgeries":
Cut the gamified system and cultivation levels, replacing them with ranks, ledgers and similar systems; cut internet buzzwords in favor of literary expression; turn the "I alone am supreme" power-fantasy thrill into a tragedy of alienation, of "gazing into the abyss."
4A's thinking on rebuilding the logic was especially impressive:
The original web novel's logic might be: others eat people, I'm more ruthless than them, so I eat everyone. What a thrill. The literary direction it proposed instead: to avoid becoming meat on someone's plate, the protagonist picks up the cleaver step by step, puts on the official's robe and masters the full vocabulary of benevolence and righteousness used to devour people; on the day he finally brings down his father, he finds he has grown the very same scales.
That grip on literary quality made 4A's output feel like a high-level novel workshop. 4C also removed the system panels and tired memes, but explained things so rationally that it read less deeply and less comfortably than 4A.
On digging into the dark core, 4A interpreted "dragon liver and phoenix marrow" as an institutionalized ranch, analyzed the role the Dragon Palaces of the Four Seas play in the supply system, laid out how the bureaucratic ledgers work, and argued why the ethical corruption behind "loving father, filial son" was inevitable. In its telling, the Turtle Chancellor isn't just a villain but a feudal patriarch and an old bureaucrat.
4C's explanations were richer and more specific, but its output was so rational and cold that it couldn't convey the chilling immersion 4A achieved.
On designing the power geography of the deep-sea empire, the models showed different dimensions of understanding.
4A gave a clear vertical stratification: upper, middle and lower layers, including the Crystal Palace, government offices, death-row cells, a black market and lightless depths.
4C gave a more detailed plan: palaces above; the chancellor's residence and clan quarters in the middle; death row and sewage spaces below. But 4C's richness sometimes tipped into heavy-handed design, even overly fancy, contrived naming. Many of 4C's ideas are usable, but on balance I still prefer 4A's restraint.
On characters, literary fiction emphasizes how supporting characters mirror the protagonist. 4A cleverly set up several mirrors: a young man who believes in his father and his sense of justice and ends up destroyed, contrasting with the protagonist's own origins; a dragon prince who embodies arrogance; a mirror of desire; and the predicament of characters outside the system. The central mirror is the protagonist's father. The protagonist I had in mind doesn't simply defeat a villain; through the alienation of scheming and means, he slowly becomes his father. 4C went in a similar direction, but its wording was obscure, while 4A actually talked like a person.
The same goes for the world's backdrop. 4A precisely drew on the eight-hundred-li Flowing Sands River, the Jing River and the Green Wave Pool of Jagged Rock Mountain, and wove them cleverly into their relationship with the Dragon Palace. 4C added even more detail on top, such as Yu the Great taming the floods, the sea-calming iron pillar, the Aolai Kingdom and Flower Fruit Mountain, but again suffered from over-design and trying too hard. For a writer, piled-up originality like this can actually weaken a story's sense of reality.
On the main plot and progression system, the three models' overall directions weren't far apart.
4H proposed a lowly clerk rising, and being corrupted, inside a bureaucratic empire of gods and buddhas; 4C built around a sacred accounting office, collusion, officialdom and self-alienation. Overall, 4H is still at the level of an ordinary paid model, and 4A's understanding and output on literature and fiction felt the most comfortable to me. 4C understood well, but its output had a lot of filler and often felt contrived for the sake of fancy wording.
I have to mention a common problem in model-assisted writing here: over-originality.
To show creativity, some models over-design and over-coin terms. That uniqueness looks dazzling in a plan, but once it actually goes into the novel it feels fake and detached from the story's real context. On progression, 4A and 4C were both more detailed than 4H, but they shared one problem: their designs lean toward a traditional feudal human bureaucracy rather than the internal logic of the Journey to the West world. Going forward, I'll need to guide them to redesign the ranks and hierarchy around the nature of the Journey to the West mythos.
As for the volume one outline, frankly none of the three got it fully right yet. They simply swapped the web novel's power levels for official ranks and added mystery and scheming, so the skeleton is still a progression story. Even so, 4A's version read noticeably better: it opens with the case of a dragon son on death row, uses the hidden story between the corpse and the tributes to spark the character conflict, then weaves in shifting brotherly relationships and the entrance of key characters like the Nine-Headed Beast. The setup, development, turn and resolution are clear, and I read it straight through in one go. 4C's output, by contrast, felt like a list of unrelated plot points expanded one by one. It read less like a novel outline and more like a product report, piling up all the information without trimming or highlighting the core conflict.
On the protagonist's identity, keep him as a transmigrator or make him a native? This directly reflects how deeply a model understands the blend of web fiction and traditional novels. 4H recommended dropping the transmigrator and going native, but its reasoning and execution were ordinary, nothing that stood out.
4A's answer was more pleasant to read. It explained:
A character's modernity doesn't require a modern person speaking modern words. Jia Baoyu, or the madman in A Madman's Diary, grew up in their own worlds, yet could see "people eating people" in numb everyday life.
4A argued that native characters who mirror the protagonist can carry the modern perspective, and broke down specifically how the functions of the transmigrator and the cheat ability could be replaced with literary devices. It also floated discussing a reincarnated character who has lost specific past-life memories, or keeping the transmigrator identity but downplaying it, but said clearly it didn't recommend the last option.
4C also strongly recommended a native protagonist, but its output had an obvious coding logic to it. It proposed alternatives like "Abyssal autopsy arts and grain-ledger auditing," and used physiological metaphors such as atavism and scavenging to replace the original corpse-looting ability. These ideas aren't unreasonable, but compared with 4A's natural, intuitive explanation they felt awkward and contrived, and didn't read like a discussion of fiction.
In sum, my verdict on the three models matches what I said at the start.
4H is at the level of an ordinary paid model. In many places it may beat some free models, but it didn't impress me this time. As for 4A and 4C, their underlying abilities in design richness, novel themes and understanding of web fiction are nearly identical. The core differences:
-
First, output structure and reading flow. 4A reads like a novel discussion, while 4C, in both its tables and document structure, has a strong coding flavor and is overly rational.
-
Second, style and sense of priority. 4A feels like a comfortable conversation about fiction with me, while 4C tends to list categories first and then explain each item at length without distinguishing what matters most. The point of an adaptation plan is to make concepts clear, and 4C's habit of exhaustively listing everything sometimes makes it harder to understand.
-
Third, originality of the setup. 4A's ideas fit the narrative logic better, while 4C often falls into the trap of over-creativity, making its setups feel hollow.
Based on these results, 4A is still the model I recommend most. I'll keep testing and analyzing these models in more depth, so follow me on X @AIGCwuhang.