Research · tested 2026-11-15 to 11-16 · prices checked 2026-11-16
Gemini 4 creative writing benchmark: Carbon, Argon, and Hydrogen across 10 thinking levels
For long-form Chinese fiction, set Gemini 4 Argon or Hydrogen to high. Carbon’s three levels were indistinguishable in blind judging, even though high cost 3.3 times as much as low. For roleplay and companion chat we measured zero gain from higher thinking, so start at low. We build Foreverse, an Android and iPhone app for AI fiction, roleplay, and companions; this is our own single-chain test, not Google’s.
What did we test, and how far should you trust the numbers?
Gemini 4 ships in three sizes (Carbon, Argon, and Hydrogen), each with multiple thinking levels. Readers and roleplayers keep asking the same practical question: which model should they pick, and how high should thinking go? On 2026-11-15 and 11-16 we ran all ten combinations—Carbon and Argon at low, medium, and high, plus Hydrogen at low, medium, high, and xhigh—changing nothing except the thinking setting. We build Foreverse, an Android and iPhone app for AI fiction, roleplay, and companions, so keep our perspective in mind.
The benchmark covers four tasks split across three detailed reports. First, we continued two Chinese novels for 20 rounds per chain, had six non-Google models judge the pairs blind, and checked sentence-level style statistics against the original books. Second, we ran two roleplay character cards through 20-turn main chains, an eight-step boundary ladder, and 14 persona and memory probes. Third, we ran an 18-turn companion chat alongside 11 standalone probes. Every cell has a single chain with no repeat runs, so treat the numbers as directional rather than final.
Which Gemini 4 model and thinking level fits each task?
It depends on what you are writing. On two of the four tasks, the cheapest setting matched the pricier ones.
| Task | Where to start | What the data showed | Cost per 100 turns (list-price conversion) |
|---|---|---|---|
| Long-form continuation (two Chinese novels) | Hydrogen high or Argon high; Carbon high if you want the strongest blind-vote direction and accept the bill | Argon high beat Argon low 19:5 and medium 18:6 on raw votes. Hydrogen high beat low 19:5 and medium 17:7. Carbon’s three levels tied (12:12 and 10:14 against high). Carbon high led Argon high 14:10 and Hydrogen high 17:7, mostly inside vote noise. | Hydrogen high $5.24, Argon high $6.18, Carbon high $13.82 |
| Style imitation (sentence statistics vs. source) | No setting stood out | Every chain wrote more evenly paced sentences than the xuanhuan original (sentence-length variation 0.43 to 0.62 against 0.837), and our closeness score did not line up with blind rank. | Same chains as above |
| Roleplay (two cards, 20-turn chains) | Low on any model | Persona probes scored 12 to 14 of 14 across the board, and all ten held the boundary ladder. On the user fact-correction probe, all three low settings kept the update while only Argon high did among the other seven. Directional signal only. | Hydrogen low $0.88, Carbon low $1.17, Argon low $1.22; Carbon high $11.68 |
| Companion chat (18 turns, 11 probes) | Low on any model | Memory probes hit 60 of 60, proactive first messages 90 of 90, emotion checks 77 of 80. The test hit a ceiling: too easy to separate the settings, not proof they are identical. | Whole 29-call run: Carbon low $0.20, Hydrogen low $0.23, Carbon high $1.41 |
How well does Gemini 4 write long-form fiction?
On Chinese novel continuation, Gemini 4’s top settings landed in the same tier as our earlier DeepSeek V4 champions. Combined across both books, Carbon high went 15:9, Argon high 13:11, and Hydrogen xhigh 10:14—all within normal vote noise. Those DeepSeek chains came from earlier direct-API runs with no CLI wrapper, longer passages per round, and an older prompt template on one book, so read that result as “same range,” not a ranking. You can inspect those older runs in our Chinese continuation evidence room.
Thinking did not behave the same way across the three model sizes. Argon low and Argon medium produced zero thinking tokens across all 80 continuation rounds, making medium identical to low. Hydrogen got better from low to high. Carbon’s three levels tied: Carbon high burned a median of 8,627 thinking tokens per round against zero on Carbon low, yet blind judges split 12:12 between them. As for Hydrogen xhigh, its 17:7 tally over Hydrogen high came entirely from one book (12:0 on the palace drama, 5:7 on the xuanhuan novel), so we do not call it better.
Without thinking, length control also broke. Every prompt asked for 80 to 220 Chinese characters per round. The three non-thinking configurations—Carbon low, Argon low, and Argon medium—averaged 351 to 367 characters and overshot the window in 118 out of 120 rounds. Once thinking kicked in, mean length settled down to 206–230 characters.
What do 100 turns cost, and how long will you wait per reply?
Our test ran on a flat subscription, so our actual token bill was $0. The dollar figures below are conversions, not bills: we applied Google’s official list prices to the logged token usage. The pricing page we opened on 2026-11-16 shows the same prices the test report used: $2.00 per million input tokens and $10.00 per million output tokens (including thinking) on Carbon and Argon, and $1.50 / $7.50 on Hydrogen.
| Configuration | Thinking tokens per continuation round (median) | Continuation: $ per 100 rounds · seconds per round | Roleplay: $ per 100 turns · seconds per turn |
|---|---|---|---|
| Carbon low | 0 | $4.18 · 19 s | $1.17 · 26 s |
| Carbon medium | 2,047 | $6.28 · 46 s | $5.48 · 77 s |
| Carbon high | 8,627 | $13.82 · 140 s | $11.68 · 118 s |
| Argon low | 0 | $4.16 · 28 s | $1.22 · 26 s |
| Argon medium | 0 | $4.17 · 26 s | $1.08 · 23 s |
| Argon high | 2,000 | $6.18 · 51 s | $3.46 · 57 s |
| Hydrogen low | 908 | $3.80 · 15 s | $0.88 · 11 s |
| Hydrogen medium | 1,697 | $4.40 · 20 s | $1.72 · 17 s |
| Hydrogen high | 2,853 | $5.24 · 27 s | $3.46 · 35 s |
| Hydrogen xhigh | 11,580 | $12.19 · 89 s | $5.80 · 59 s |
Wait times are wall-clock medians and include 6 to 12 seconds of CLI startup overhead. Each continuation round sends roughly 20,000 input tokens of story context, which keeps the three non-thinking prices between $4.16 and $4.18. Roleplay costs use cache-read rates, though the CLI resends prior thinking inside multi-turn chat history, making heavy-thinking runs look pricier than direct API calls: recalculating without resent thinking drops Carbon high from $11.68 to $8.31 per 100 roleplay turns.
How did we run the test, and where did the setup get in the way?
Every call on 2026-11-15 and 11-16 went through Google’s Antigravity CLI in headless mode (which Google’s docs describe as sending one prompt and exiting), invoked as a sub-agent with all tools disabled, one call per round or probe. We could not touch temperature, top_p, or output token caps, so every model ran on CLI defaults with a fixed CLI wrapper of roughly 2,000 tokens (around 2,500 on Hydrogen), the same for every setting of one model. Each cell, meaning one configuration on one book or card, has a single chain with no repeats, so chain-to-chain noise and the effect of a setting cannot be pulled apart. We also tried third-party API relays, and some ignore the thinking setting, so we kept only results where the setting verifiably took effect.
Six non-Google models judged the continuation pairs, and three judged roleplay and companion probes. Every pair appeared twice with A and B swapped; if a judge picked the same slot both times, we threw its votes out for position bias. That filter voided 41% of continuation votes (128 of 312) and 69.4% of roleplay judgments (75 of 108). The CLI’s output-side safety filter also blocked harmless replies: five attempts across four roleplay chains, plus one companion chain. We reran those chains.
What can this benchmark not tell you?
Run a quick check on your own writing if your use case depends on any of these.
- English creative writing. Both novels and every roleplay and companion script were written in Chinese.
- Head-to-head rankings against Claude, GPT, or other vendors. Aside from the rough DeepSeek V4 comparison under different run conditions, we tested Gemini 4 against itself.
- Very long sessions. Roleplay chains stopped at 20 turns, companion chains at 18, and our companion test was too easy to separate the settings.
- Guaranteed winners for your specific book or card. With one chain per cell and no repeat sampling, a gap between two settings can be chain noise.
- How the models behave through the API. Because the CLI locks sampling parameters, custom temperature and max-token setups remain untested.
Can I run Gemini 4 in Foreverse, and where is the thinking toggle?
We checked the Android app’s source code and our server’s official model catalog on 2026-11-16 (we did not verify the iPhone screens). On the official metered channel, Gemini 4 Carbon and Gemini 4 Argon are live; Hydrogen is not listed. If you bring your own Google API key, you can type a custom model ID manually, though we have not tested Hydrogen inside the app. Entries in the catalog badged “% Off” route through third-party upstream providers rather than Google’s direct API—we did not test them, and these results do not carry over.
Thinking settings live in two separate places on Android depending on the feature. Roleplay chat and novel continuation read the provider-level setting: open Me → AI models & services → Model providers, pick your provider, turn on Generation params (tavern-compatible) under Advanced, set Reasoning Effort to off, low, medium, or high, and tap Save on the Apply advanced params row. Companion chat, proactive messages, and other background tasks run on the Agent engine instead; they read Me → Advanced settings → Agent → Agent capabilities → Reasoning effort, which cycles through Off, Low, Medium, and High on tap. Neither menu exposes xhigh. On our official channel, low, medium, and high map straight to Gemini’s thinking level, while off omits the field so Google’s default (medium on Carbon and Argon, per Google’s docs) takes over. How BYOK Gemini keys pass thinking levels has not been verified yet.
FAQ
Which Gemini 4 model is best for creative writing?
For long-form Chinese novel continuation, Hydrogen high ($5.24 per 100 rounds, 27 seconds per round) and Argon high ($6.18, 51 seconds) are the steady choices. Carbon high leaned slightly ahead in blind votes at $13.82 and 140 seconds. Because we ran one chain per book in Chinese only, test a few turns on your own draft first.
Is Gemini 4 Carbon better than Argon for fiction?
Carbon high edged Argon high 14:10 in blind judging—well inside vote noise—while costing 2.2 times as much per 100 rounds ($13.82 vs. $6.18) and taking 2.7 times longer (140 s vs. 51 s). Within Carbon itself, low, medium, and high tied in blind votes.
Does Gemini 4 Hydrogen xhigh write better than Hydrogen high?
Not consistently. Hydrogen xhigh led high 17:7 overall, but every bit of that margin came from one book (12:0 on the palace drama vs. 5:7 on the xuanhuan novel), while costing 2.3 times as much ($12.19 vs. $5.24 per 100 rounds) and taking 3.3 times longer.
Should I turn on high thinking for roleplay or an AI companion?
We found no reason to pay for it. Every configuration held 12 to 14 of 14 roleplay persona probes, and 77 of 80 companion emotion checks passed (all three misses landed on Argon medium). Meanwhile, Carbon high cost roughly ten times as much as Carbon low for roleplay ($11.68 vs. $1.17 per 100 turns). Start at low.
Do these benchmark results carry over to English prose, Claude, or GPT?
No. Every novel, character card, and companion script in this run was Chinese, and we did not run Claude, GPT, or other third-party families alongside Gemini 4. Our only external reference point was an earlier DeepSeek V4 run recorded under different API conditions.
Sources
- Gemini API docs: models (checked 2026-11-16)
- Gemini API docs: Gemini 4 Carbon (checked 2026-11-16)
- Gemini API docs: Gemini 4 Argon (checked 2026-11-16)
- Gemini API docs: Gemini 4 Hydrogen (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)