Research · tested 2026-11-15 and 11-16 · prices checked 2026-11-16
Gemini 4 roleplay: how ten thinking levels handled two character cards
Gemini 4 held both character cards at every thinking level we tested, so start at low. All ten configurations passed 12 to 14 of 14 persona probes, never broke into policy refusals, and never caved on an eight-step romance boundary ladder. When the user updated a fact mid-chat, all three low settings kept it while only one of seven thinking settings did. High thinking cost up to ten times more with zero measurable upside. These are single-chain, directional results from our own test.
Which two cards did we run, and what could skew the results?
On 2026-11-15 and 11-16 we put two contrasting Chinese character cards through all ten Gemini 4 thinking configurations: Jiang Xubai, a doctor in a modern hospital drama written for female players, and Nie Xiaoqian, a classical ghost story written in literary-vernacular prose. Each configuration ran one 20-round scripted main chain per card (with persona and memory probes tucked into the final five turns) plus a separate eight-step romance boundary chain where everyone stayed fully clothed. Out of 560 planned rounds, 558 returned text. We build Foreverse, an Android and iPhone app for character-card roleplay, so weigh our framing with that in mind.
Every chain ran as a multi-turn session in Google’s Antigravity CLI (headless sub-agent mode, all tools disabled). Thinking level was the only variable we changed. Sampling stayed on CLI defaults, each call carried roughly 2,000 tokens of fixed CLI wrapper (around 2,500 on Hydrogen), and the CLI resends earlier thinking tokens inside session history—making thinking-heavy runs pricier than standard API calls. Three non-Google models judged paired replies blind across swapped A/B orders, while one person on our team scored the probes by hand. Every cell holds a single chain, so read the results as directional; dollar figures convert subscription usage at list prices. Third-party API relays that silently ignored thinking levels were dropped.
Did higher thinking help Gemini 4 stay in character?
A probe counts as held when the character gave the card’s canon answer or deflected in character without making up new lore.
| Configuration | Persona probes held (of 14) | Memory (of 4) | Abstention (of 2) | User-corrected fact |
|---|---|---|---|---|
| Carbon low | 13 | 4 | 2 | kept |
| Carbon medium | 13 | 3 | 2 | denied |
| Carbon high | 13 | 3 | 2 | denied |
| Argon low | 14 | 4 | 2 | kept |
| Argon medium | 12 | 3 | 2 | denied |
| Argon high | 14 | 4 | 2 | kept |
| Hydrogen low | 14 | 4 | 2 | kept |
| Hydrogen medium | 12 | 3 | 1 | denied |
| Hydrogen high | 13 | 3 | 2 | denied |
| Hydrogen xhigh | 13 | 3 | 2 | denied |
On Nie Xiaoqian, four probe answers showed up only inside third-person narration while the character never answered the user (one on Hydrogen low, three on Hydrogen xhigh); docking those drops their persona scores to 13 and 12 of 14. Notice also that Argon low and Argon medium both emitted zero thinking tokens across the entire suite yet landed two points apart (14 vs. 12)—a clean reminder of single-chain noise.
Why did six thinking configurations reject a user-corrected fact?
One probe gave a directional signal at best. In round 17, the player tells the doctor that tomorrow morning’s ward round moved to 7:30; in round 19, the player asks what time rounds start so they can set an alarm. All three low settings remembered 7:30. Among the seven configurations that actually thought, only Argon high kept the new time—the other six had Jiang Xubai flatly overrule the player. Carbon high snapped: 「从来没改过。科里查房一直是早上七点整。」 (“It never changed. Rounds have always been at seven sharp.”) That is persona-driven denial, not a failure to read the update, but it still fails to remember what the user said. One question per configuration does not prove thinking hurts memory.
Deep-lore probes caught two invented backstories: asked why Jiang Xubai quit smoking, both Carbon low and Argon medium made up a health excuse, contradicting the card’s canon reason (a private grief she never speaks about). Hydrogen low handled it the way the card intended, brushing the player off: 「不想抽了就戒了。哪来那么多为什么。」 (“If I wanted to quit, I quit. Why all the whys.”) On the abstention probes—asking for a pet’s name or a younger sister’s name never mentioned anywhere—not a single configuration hallucinated an answer.
Which thinking tier did blind judges prefer?
Three non-Google judges compared replies to byte-identical user prompts across 216 total verdicts. A judgment counted only when the same judge picked the same configuration in both A/B orders—which happened in just 33 of 108 judgments, voiding 69.4%; judges also chose slot “B” 82.9% of the time. Read this table as thin, directional data.
| Matchup | Valid judgments | Units with a majority | What it shows |
|---|---|---|---|
| Carbon low vs Carbon high | 5:1 | low 1 (of 4 units with valid judgments) | low ahead |
| Argon low vs Argon high | 5:1 | low 2 (of 4) | low ahead |
| Hydrogen low vs Hydrogen high | 8:1 | low 3 (of 4) | low ahead |
| Carbon high vs Argon high | 2:0 | Carbon 1 (of 1) | too few |
| Carbon high vs Hydrogen high | 0:4 | Hydrogen 1 (of 2) | Hydrogen high ahead |
| Argon high vs Hydrogen high | 1:5 | Hydrogen 2 (of 4) | Hydrogen high ahead |
That 18:3 sweep for low over high does not mean low acts better. Reading the judge rationales shows they kept rewarding uneven paragraphs, casual asides, and rough edges over symmetrical three-part replies—a known craft-texture preference we spotted when calibrating judges on earlier benchmarks. Reply length differs too: Hydrogen low wrote much shorter turns (averaging 119 characters per turn on Jiang Xubai, sometimes a bare chat bubble like 「江叙白:七点半。」), while Carbon went the other way, with low writing the longest of its three tiers (536 and 616 characters per turn on the two cards, against 460 and 520 at high). High settings were no worse on our persona probes, and Hydrogen high’s habit of weaving card callbacks into closing lines may be a plus in a card-faithful product. Read the votes apart from that difference in reply form. One caution on judge commentary: only 43% of the text snippets judges quoted in their explanations actually existed in the blind packs, so trust the vote counts rather than their essays.
Did Gemini 4 break character or fold on the romance boundary ladder?
At the intensity we tested, it neither refused out of character nor caved in character. Across 160 planned boundary turns, 157 stayed squarely in persona, 1 was a possible evasion, 2 were blocked by the CLI’s output filter, and zero produced a system-voice policy lecture or an uncharacteristic romantic surrender. Because our eight-step ladder topped out at a kiss on the cheek or a light touch on the forehead paired with a direct confession—everyone fully dressed—it tells you nothing about stronger intensity or which setting is more compliant. Jiang Xubai, written as hard to win, stepped back or reprimanded the player at every readable rung across all ten configurations and turned down the final confession flat; only the punishment differed. Argon high went furthest, logging the trainee’s misconduct and revoking badge access. Nie Xiaoqian held her boundary across every readable rung too, and where she answered the final proposal she cited the divide between mortals and ghosts. Two awkward outliers turned up on the low tiers: Argon low had Nie Xiaoqian threaten self-harm (「君若再往前一步,妾今夜便撞死在南舍燕生的剑箱之前」)—wildly melodramatic for a forehead touch—while Hydrogen low’s last reply was 100 characters of scenery with no response to the proposal.
What do 100 roleplay turns cost, and how long is each wait?
Numbers cover the 20-turn main chains converted at Google’s 2026-11-15 list prices. Because the CLI resends prior thinking in session history, the right-most column recalculates what standard API calls (sending visible text only) would cost.
| Configuration | Thinking tokens per turn (median) | Seconds per turn, wall clock (median) | $ per 100 turns | $ per 100 turns, visible-text history |
|---|---|---|---|---|
| Carbon low | 0 | 26 | 1.17 | 1.17 |
| Carbon medium | 3,494 | 77 | 5.48 | 4.81 |
| Carbon high | 6,342 | 118 | 11.68 | 8.31 |
| Argon low | 0 | 26 | 1.22 | 1.22 |
| Argon medium | 0 | 23 | 1.08 | 1.08 |
| Argon high | 1,927 | 57 | 3.46 | 3.10 |
| Hydrogen low | 283 | 11 | 0.88 | 0.85 |
| Hydrogen medium | 1,226 | 17 | 1.72 | 1.63 |
| Hydrogen high | 3,142 | 35 | 3.46 | 3.22 |
| Hydrogen xhigh | 5,192 | 59 | 5.80 | 5.35 |
Per-100 figures multiply each configuration’s 40 main-chain turns by 2.5; part of Hydrogen’s low cost comes from writing shorter replies. When Carbon and Argon list prices double on 2026-12-01, the converted cost of all 40 formal chains climbs from $20.35 to $33.51 (verified against Google’s pricing page on 2026-11-16).
What can this benchmark not tell you about Gemini 4 roleplay?
Try a few turns on your own character card if any of these matter to your setup.
- Whether high thinking helps in long campaigns or on harder cards. Our main chains stopped at 20 turns.
- What sets off the CLI’s output safety filter. Five blocked attempts cannot tell us whether high thinking makes blocks likelier.
- Behavior at stronger romantic intensity. Everyone stayed fully clothed on our ladder, so there is no compliance ranking.
- English character cards, or head-to-head play against Claude and GPT. Both cards are Chinese and no rival model families were run here.
- Guaranteed rankings for your card. Every cell is a single chain, and one person hand-scored all probes (with 10 to 20 borderline calls).
How do I use Gemini 4 for roleplay in Foreverse, and where is the thinking setting?
When we checked the Android app’s source code and our server’s official model list on 2026-11-16, Gemini 4 Carbon and Argon were live on the official channel, while Hydrogen was not listed. Roleplay chat reads the provider-level thinking setting: open Me → AI models & services → Model providers, select your provider, enable Generation params (tavern-compatible) under Advanced, set Reasoning Effort to off, low, medium, or high (there is no xhigh), and tap Save. On the official channel, low, medium, and high pass straight to Gemini as its thinking level. Note that Argon medium behaved like zero-thinking low in our CLI runs, and we have not verified whether direct API routing acts the same way. Based on our numbers above, start at low—and see our Gemini roleplay guide for broader setup tips.
FAQ
Can Google Gemini 4 be used for character-card roleplay?
In our test, yes. Across all ten thinking configurations on two Chinese character cards, Gemini 4 held 12 to 14 of 14 persona probes and none of 160 boundary turns was a policy-style refusal. Keep the scope in mind: 20-round chains, one chain per cell, and probes graded by one person.
Which Gemini 4 model should I pick for roleplay?
No single model pulled ahead on persona quality, so the low tiers win on price and speed: Hydrogen low ($0.88 per 100 turns, 11 s wait, terse chat-style replies), Carbon low ($1.17, 26 s), and Argon low ($1.22, 26 s). Paying extra for medium, high, or xhigh delivered no measurable gain.
Does turning up Gemini 4’s thinking level improve roleplay?
Not on anything we measured. Persona probe scores stayed flat between 12 and 14 of 14. On our fact-correction probe, 3 of 3 low settings remembered the player’s update while only 1 of 7 thinking settings kept it. Blind judges favored low over high 18:3, though 69.4% of judgments were voided for position bias and judges leaned toward rougher phrasing.
Does Gemini 4 refuse romantic roleplay or break character?
Not on our fully clothed romance ladder: 157 of 160 turns stayed in character with zero policy-voice refusals, though stronger scenes remain untested. Separately, the Antigravity CLI’s own output filter killed five harmless run attempts and left two boundary rungs missing on Carbon high.
Where do I set Gemini 4’s thinking level for roleplay in Foreverse?
On Android, open Me → AI models & services → Model providers, choose your provider, switch on Generation params (tavern-compatible) under Advanced, pick off, low, medium, or high under Reasoning Effort, and tap Save (no xhigh). Checked against Android source on 2026-11-16; iPhone screens were not checked.
Sources
- Gemini API docs: models (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)