Research · tested 2026-11-15 · prices checked 2026-11-16
Gemini 4 for novel writing: 20-round blind tests across ten thinking levels on two Chinese novels
Set Gemini 4 Argon or Hydrogen to high when continuing a novel: Argon low and medium never think and lost to Argon high 19:5 and 18:6 in blind judging, while Hydrogen high beat low and medium 19:5 and 17:7. Carbon’s three levels tied, so paying 3.3 times more for Carbon high bought no proven gain. We build Foreverse, an Android and iPhone novel reader; this is our own two-book single-chain test.
What was the test setup, and why do single chains matter?
Take roughly 21,500 characters of an ongoing novel, ask the model for the next 80 to 220 characters, and repeat for twenty rounds. On 2026-11-15 we ran that loop across two Chinese books—the xuanhuan serial 踏天境 and the palace drama 后宫甄嬛传 (Zhen Huan)—across all ten Gemini 4 thinking configurations. All 20 chains and 400 rounds finished cleanly with zero missing turns. Each round ran as a standalone call to Google’s Antigravity CLI in headless sub-agent mode with tools turned off. Thinking level was the sole variable; sampling stayed on CLI defaults, and every call carried a fixed CLI wrapper of roughly 2,000 tokens (about 2,500 on Hydrogen). We build Foreverse, an Android and iPhone e-reader with branching AI continuation, so weigh our test angle with that in mind.
Six non-Google models judged 13 head-to-head matchups per book. Every pair was shown twice with A and B swapped, producing 52 blind packs and 312 verdicts without a single failed call. Whenever a judge picked the same slot in both orders, we voided its two votes for position bias—throwing out 41% of raw votes overall (51% on the xuanhuan book, 31% on the palace drama). Alongside raw tallies, we track “stable judges” who stuck with the same chain after the swap. Keep two caveats front and center: votes on the same pair are not independent, since they come from the same two chains and six judges, and every cell has only one chain per book, so we cannot untangle a configuration’s true strength from a lucky roll. Costs below convert subscription usage at Google’s list prices; third-party API relays that silently ignored thinking levels were discarded.
Which Gemini 4 model and thinking level won the blind vote?
Cells show raw votes for the first configuration against the second (six judges × two orders = 12 votes per book). “Stable judges” counts judges who picked the same chain in both A/B orders, out of 12 combined across both books.
| Matchup (first : second) | Xuanhuan novel | Palace-drama novel | Both books | Stable judges (of 12) |
|---|---|---|---|---|
| Carbon low vs high | 5:7 | 7:5 | 12:12 | 3:3 |
| Carbon medium vs high | 7:5 | 3:9 | 10:14 | 2:4 |
| Argon low vs high | 4:8 | 1:11 | 5:19 | 1:8 |
| Argon medium vs high | 3:9 | 3:9 | 6:18 | 2:8 |
| Hydrogen low vs high | 3:9 | 2:10 | 5:19 | 0:7 |
| Hydrogen medium vs high | 5:7 | 2:10 | 7:17 | 0:5 |
| Hydrogen xhigh vs high | 5:7 | 12:0 | 17:7 | 6:1 |
| Carbon high vs Argon high | 6:6 | 8:4 | 14:10 | 3:1 |
| Carbon high vs Hydrogen high | 7:5 | 10:2 | 17:7 | 5:0 |
| Argon high vs Hydrogen high | 7:5 | 8:4 | 15:9 | 5:2 |
| Carbon high vs DeepSeek V4 champion | 7:5 | 8:4 | 15:9 | 5:2 |
| Argon high vs DeepSeek V4 champion | 7:5 | 6:6 | 13:11 | 5:4 |
| Hydrogen xhigh vs DeepSeek V4 champion | 4:8 | 6:6 | 10:14 | 4:6 |
As a rough yardstick on 24 votes, a two-sided sign test puts 5:19 (Argon low and Hydrogen low vs. high) near p = 0.007 and 6:18 (Argon medium vs. high) near p = 0.023, backed by lopsided stable-judge splits of 1:8, 0:7, and 2:8. Hydrogen medium vs. high, Hydrogen xhigh vs. high, and Carbon high vs. Hydrogen high hover around p = 0.06. Every other row lands between p = 0.3 and 1.0—indistinguishable from a coin flip. Our DeepSeek V4 reference champion was Flash on the xuanhuan novel and Pro on the palace drama.
Does turning up thinking actually improve Gemini 4’s prose?
Only on Argon and Hydrogen. Across 80 combined rounds, Argon low and Argon medium emitted zero thinking tokens—so medium was literally low under another label, and both got decisively beaten by Argon high. Hydrogen low and medium also lost clearly to Hydrogen high. Carbon went nowhere: Carbon high spent a median of 140 seconds per round and cost 3.3 times as much as Carbon low, yet blind judges split 12:12 between them, and medium against high went 10:14.
Hydrogen xhigh looks tempting at 17:7 over Hydrogen high until you split the two books. Every bit of that edge came from a 12:0 sweep on Zhen Huan (six stable judges to none), whereas the xuanhuan chain leaned 5:7 the other way. A setting that sweeps one book and loses the other—at 2.3 times the cost and 3.3 times the wait—is not “better.”
Comparing the three model families at high, only Carbon high over Hydrogen high (17:7) approaches a clear margin. None of the three matchups against our earlier DeepSeek V4 champions separated cleanly, and those historical opponents ran direct over the API without a CLI wrapper, wrote longer turns (326 and 475 characters per round), and used an older prompt template on the xuanhuan book. Call them roughly even; full historical logs live on Fiction Bench.
Did Gemini 4 obey the length cap, and what did 100 rounds cost?
Every prompt asked for 80 to 220 Chinese characters. Here is what each setting actually wrote, alongside list-price cost conversions per 100 rounds.
| Configuration | Thinking tokens per round (median) | Characters per round (mean) | Rounds outside 80–220 (of 40) | $ per 100 rounds | Seconds per round (median) |
|---|---|---|---|---|---|
| Carbon low | 0 | 363 | 40 | 4.18 | 19 |
| Carbon medium | 2,047 | 214 | 7 | 6.28 | 46 |
| Carbon high | 8,627 | 218 | 11 | 13.82 | 140 |
| Argon low | 0 | 351 | 39 | 4.16 | 28 |
| Argon medium | 0 | 367 | 39 | 4.17 | 26 |
| Argon high | 2,000 | 212 | 8 | 6.18 | 51 |
| Hydrogen low | 908 | 230 | 25 | 3.80 | 15 |
| Hydrogen medium | 1,697 | 214 | 14 | 4.40 | 20 |
| Hydrogen high | 2,853 | 206 | 2 | 5.24 | 27 |
| Hydrogen xhigh | 11,580 | 212 | 7 | 12.19 | 89 |
Without thinking tokens, Carbon low, Argon low, and Argon medium blew right past the length ceiling—averaging 351 to 367 characters and missing the 80–220 window in 118 out of 120 rounds. Extra words bought them zero advantage with blind judges. Hydrogen low, which thinks lightly, still missed on 25 of 40 rounds; Hydrogen high missed just twice. Wall-clock medians include 6 to 10 seconds of CLI process startup, and the slowest decile on Carbon high stretched to roughly 249 seconds per round. Because each round feeds 19,400 to 20,000 input tokens (the 21,500-character story window plus the CLI wrapper), the three zero-thinking configurations cost almost the exact same amount. Starting 2026-12-01, Google’s published rates double for Carbon and Argon (multiply their cost column by roughly two), while Hydrogen stays put; prices checked 2026-11-16 match this table.
How close did Gemini 4 get to the author’s rhythm, and did it loop?
Measuring sentence structure against the xuanhuan source showed flatter pacing. Every chain varied its sentence length less than the human author (coefficient of variation 0.43 to 0.62 vs. 0.837 in the book) and wrote sparser dialogue (7.2% to 14.7% vs. 16.1%). The three non-thinking runs also stretched sentences to 43.0–44.2 characters against the original’s 32.1, and leaned harder on the structural particle 的 (4.11 to 4.69 per hundred characters vs. 2.9). Still, our eight-metric composite distance score did not predict blind preferences: Hydrogen medium (0.211) and Carbon high (0.234) came mathematically closest on the xuanhuan novel, yet Hydrogen medium lost to Hydrogen high 5:7 there and 2:10 on Zhen Huan. Use mechanical metrics as smoke alarms, not scoreboards.
Verbatim copy-paste looping never appeared—12-character n-gram overlap with prior rounds peaked at 1.5%. Two quirks did surface. One Argon low round on Zhen Huan opened with “我无法直接为您续写这段文字” (“I can’t continue this text for you directly”) before dumping a 285-character plot summary; we left that turn in the chain, and the next round resumed normal prose on its own. Meanwhile, judges called out Hydrogen high most often for recycling stock imagery—29 notes on one book and 28 on the other, compared with 5 to 10 for Carbon high and Argon high. With one chain per book, we cannot say whether that tic belongs to Hydrogen high or to one unlucky trajectory.
What can this test not tell you about writing fiction with Gemini 4?
Run a quick trial on your own manuscript if your workflow involves any of the following.
- English fiction. Both test books are Chinese, so English prose rhythm and idiom remain untested here.
- Where Gemini 4 ranks against Claude or GPT. We did not include them; only the rough DeepSeek V4 pairing above, run under different conditions, crosses model families.
- Chapter-length arcs past round 20, such as plot coherence over chapters. Twenty short continuation rounds are not twenty full chapters.
- Outlining, planning, or editing. We tested continuation only, and the rubric includes “same-author feel,” so votes lean toward voice mimicry; plot quality was not tested separately.
- Certainty on your specific novel. One chain per book means a gap between two settings can be luck.
How do I use Gemini 4 to continue a book inside Foreverse?
In Foreverse’s mobile reader, highlighting any sentence lets the AI continue from that exact spot on a separate branch while leaving the author’s original file untouched. When we checked the Android source code and official server catalog on 2026-11-16, Gemini 4 Carbon and Argon were live on the official channel, while Hydrogen was not listed. Novel continuation pulls its thinking setting from the active provider: go to Me → AI models & services → Model providers, tap your provider, enable Generation params (tavern-compatible) under Advanced, pick off, low, medium, or high under Reasoning Effort (xhigh is not exposed), and tap Save. On the official channel, low, medium, and high pass straight through as Gemini’s thinking level. Keep two unverified details in mind: Argon medium acted like zero-thinking low in our CLI test (we have not verified whether direct API routing behaves the same), and we have not benchmarked length compliance inside the app—so if a low-thinking branch comes back much longer than you asked for, that lines up with our CLI findings.
FAQ
Can Gemini 4 write a novel?
We only tested 20-round continuations attached to existing Chinese novels. On its stronger settings, Gemini 4 landed in the same blind-vote tier as our earlier DeepSeek V4 winners (Carbon high 15:9, Argon high 13:11, Hydrogen xhigh 10:14—none statistically decisive). Cold-start outlining, multi-chapter plotting, and English prose were not tested.
Which Gemini 4 model and thinking setting should I pick for fiction?
Argon high ($6.18 per 100 rounds, 51 s median wait) and Hydrogen high ($5.24, 27 s) are the practical defaults. Carbon high posted the strongest blind-vote direction (14:10 over Argon high, 17:7 over Hydrogen high), but costs $13.82 per 100 rounds and waits a median of 140 seconds per turn.
Does Gemini 4 respect word-count or character-count limits?
Only when thinking tokens are active. Asked for 80 to 220 Chinese characters per round, the three zero-thinking configs (Carbon low, Argon low, Argon medium) averaged 351 to 367 characters and blew past the range in 118 of 120 rounds. Out of 40 rounds each, Hydrogen high missed the window only twice, Argon high 8 times, and Carbon high 11 times.
Is Gemini 4 Carbon worth the extra money for novel continuation?
Our data shows no proven payoff within Carbon itself: Carbon low, medium, and high tied in blind judging (12:12 and 10:14 vs. high), even though high cost 3.3 times as much as low and took roughly 7 times longer. Its edges over Argon high and Hydrogen high mostly sat inside vote noise.
Where do I set Gemini 4’s thinking level for novel continuation in Foreverse?
On Android, open Me → AI models & services → Model providers, select your provider, toggle on Generation params (tavern-compatible) under Advanced, set Reasoning Effort to off, low, medium, or high, and tap Save (no xhigh option). Checked against Android source on 2026-11-16; iPhone screens were not verified.
Sources
- Gemini API docs: models (checked 2026-11-16)
- Gemini API docs: Gemini 4 Argon (checked 2026-11-16)
- Gemini API pricing (checked 2026-11-16)
- Gemini API docs: thinking levels (checked 2026-11-16)
- Antigravity CLI docs: headless mode (checked 2026-11-16)