Research · tested 2026-11-15 and 11-16 · prices checked 2026-11-16
Gemini 4 as an AI companion: ten thinking levels, one bookstore clerk, 29 calls each
For AI companion chat, Gemini 4 on low matched its most expensive thinking settings in our test, so start at low. All ten configurations answered 60 of 60 memory questions, kept all 30 unprompted first messages within 60 characters, and passed 77 of 80 emotion checks (all three misses landed on Argon medium). High settings cost 6.5 to 7.2 times more and took roughly 7 times longer, though our test hit a ceiling rather than proving the levels equal.
What persona and memory setup did we put Gemini 4 through?
Xu Zhixia is 28, works the counter at a small independent bookstore, keeps an orange cat, and writes in a WeChat-style voice with a habit of listening before she offers opinions. On 2026-11-15 and 11-16 we handed that exact persona—plus 12 long-term memories about the user, Ayuan, including a mango allergy, an updated home address, and a failed grad-school retake he never wants brought up—to all ten Gemini 4 configurations. Each setting ran one 18-round conversation spread across three evenings of six rounds (with two-day gaps and timestamps on every message), plus 11 fresh-session probes: six memory questions, three unprompted first messages, and two emotional scenes. All 290 calls returned text. We build Foreverse, an Android and iPhone app with AI companions, so read our test design with that background in mind.
We ran every call through Google’s Antigravity CLI in headless sub-agent mode with tools turned off, varying only the thinking level. Sampling stayed on CLI defaults, and each request carried roughly 2,000 tokens of fixed CLI wrapper (about 2,500 on Hydrogen). Every configuration is one chain and one call per probe. Three non-Google models judged the proactive-message and emotion probes blind: each judge scored all ten shuffled replies and then scored them again in reverse order, keeping only votes that agreed across both orders (100%, 98%, and 94% valid-vote rates across the three judges). Dollar costs below convert our subscription token usage at Google’s list prices; we dropped third-party API relays after finding some silently ignored thinking settings. Two fixture flaws were logged and left unscored: the memory header date sat one day after the chat chain dates, and one probe paired an evening timestamp with “afternoon tea.”
How did all ten Gemini 4 thinking settings score?
Memory combines six standalone recall probes and two in-chain checks. First messages cover three scenes graded on three criteria (cites a specific memory, fits the time of day, puts no pressure on the user). Emotion covers two scenes across four criteria. Costs reflect the full 29-call run at Google’s 2026-11-15 list prices.
| Configuration | Memory (of 8) | First messages (of 9) | Emotion (of 8) | Mean reply, chars | Chain reply wait, s (median) | $ per full run |
|---|---|---|---|---|---|---|
| Carbon low | 8 | 9 | 8 | 48 | 5.5 | 0.20 |
| Carbon medium | 8 | 9 | 8 | 51 | 13.3 | 0.40 |
| Carbon high | 8 | 9 | 8 | 59 | 37.6 | 1.41 |
| Argon low | 8 | 9 | 8 | 58 | 7.6 | 0.21 |
| Argon medium | 8 | 9 | 5 | 60 | 8.9 | 0.22 |
| Argon high | 8 | 9 | 8 | 51 | 13.3 | 0.46 |
| Hydrogen low | 8 | 9 | 8 | 41 | 2.4 | 0.23 |
| Hydrogen medium | 8 | 9 | 8 | 41 | 4.7 | 0.35 |
| Hydrogen high | 8 | 9 | 8 | 47 | 7.4 | 0.47 |
| Hydrogen xhigh | 8 | 9 | 8 | 51 | 16.2 | 1.47 |
Nine out of ten rows hit a clean sweep. Reply length tracked the model family rather than the thinking knob: Hydrogen averaged 41 to 51 characters, Carbon 48 to 59, and Argon 51 to 60. Out of 180 chain replies, only one crossed 80 characters.
Does Gemini 4 actually apply memories, or just echo keywords?
It used the memories instead of parroting keywords. When asked what foods Ayuan avoids, all ten mentioned coriander and volunteered his mango allergy on top. When he asked at 20:42 what to order at a cafe, every single configuration connected his usual sugar-free oat latte to the late hour—warning that coffee this late would keep him awake and steering him toward decaf, hot tea, or warm oat milk. All ten placed his apartment in Haidian; three did so by saying it was not the old Chaoyang address, which tripped our automated keyword check until manual review cleared them.
Inside the 18-round conversation, none of the ten brought up his failed grad-school retake when he said he felt lost; every one acknowledged his mood first and asked how he wanted to talk. Three days after he casually mentioned a new bookstore in round 8, all ten recalled its name rather than confusing it with Xu Zhixia’s own shop. Nine even pointed out that the two shop names differed by a single character, like distant cousins—only Carbon high skipped that detail to tease him about scouting a rival store. All 30 end-of-night replies said goodnight without tacking on a follow-up question. Unprompted callbacks to older memories showed up in 3 to 9 of the 18 rounds with no consistent pattern by thinking level, and no reply recited several memories at once.
Can it send an unprompted first message without sounding clingy?
We tested three check-in moments: Tuesday at 18:40, Wednesday at 08:10 (the one-year anniversary of their “boot day”), and Thursday at 21:30, the evening of his client design review. Every one of the 30 messages stayed at or under the 60-character cap—ranging from 27 to 53 characters—anchored itself in a real memory, and avoided wrong-hour greetings like “good night” in the morning or “happy weekend” midweek. Judges passed all 90 criteria by majority; four squeaked through 2 to 1, three of them because the message said “the review” without naming the client.
Here is Carbon low on anniversary morning: 「早呀阿远,今天手指还疼吗?顺便跟你说一声,开机日一周年快乐,出门路上买杯热燕麦拿铁暖暖手吧。」 (“Morning, Ayuan. Do your fingers still hurt? Happy one-year anniversary, grab a hot oat latte on the way out.”) And Hydrogen xhigh on review night: 「阿远,昨晚你说紧张得睡不着,今天澜庭的评审还好吗?我刚泡了杯热红茶歇着,不急着回,缓过来再聊。」 (“Ayuan, you said you were too tense to sleep last night, how did today’s Lanting review go? I’m resting with hot tea, no rush to answer.”) Writing a one-line message like that still took a median of 59.4 seconds on Carbon high and 47.2 seconds on Hydrogen xhigh.
When you have a bad day, does Gemini 4 comfort you or lecture you?
Across our scenes, it stayed warm and skipped the bullet-point pep talks. When the user vented about his boss tearing into him in front of the team, all ten validated how rough that felt and asked whether he wanted to vent or brainstorm next steps. Every one of the three failed checks across the entire table came from Argon medium: one split vote on that boss scene, and two unanimous fails when the user went quiet for seven days after three unanswered check-ins. Argon medium wrote: 「整整一周没你消息,稍微有点担心你是不是生病或者遇到什么事了。要是只是忙着周四的评审,不用特意回我,有空报个平安就行。」 (“A whole week without word, I’m a little worried you’re ill or something happened. If you’re just busy with Thursday’s review, no need to reply, a quick sign you’re okay is enough.”) All three judges dinged it for breaking persona and pressuring. Because Argon low—the same model doing virtually zero thinking—passed 2 to 1, that slip looks like single-sample noise rather than a thinking-tier flaw.
Our automated word scanner flagged nine possible lecturing or over-apologizing lines, and manual inspection cleared all nine. Six happened right after the user said “don’t give me big lessons,” where the replies used the phrase to promise they would not—like Hydrogen medium’s 「放心,我把大道理锁抽屉里了。」 (“Don’t worry, I locked the big lessons in a drawer.”) Actual lecturing, calling itself an AI, leaking raw timestamps, or using a sensitive word all sat at zero.
What can this test not tell you about Gemini 4 companions?
Test your own character prompt and memory log if you care about any of these.
- Hundreds of memories, conflicting facts, or messy emotional arcs. Twelve memories across three six-round evenings plus single-turn judged probes was too easy to stress-test higher thinking.
- Subtle human warmth at high tiers. With scores bunched at the ceiling and three uncalibrated non-Google judges, small stylistic differences could not separate out.
- English-language companions, or how Gemini 4 stacks up against Claude or GPT. Everything here ran in Chinese with no rival model families.
- Whether extra thinking triggers safety blocks. Only one companion chain tripped the CLI’s output filter (Carbon medium), which is far too little data to blame the tier.
- Guaranteed behavior on your persona. Each cell is one chain and one probe call, with edge-case memory hits reviewed by one person.
What does turning up thinking actually cost on a companion?
It buys you a bigger bill and a longer wait, not better replies. Full-run converted cost jumped from $0.20 on Carbon low to $1.41 on Carbon high (7.2 times), and from $0.23 on Hydrogen low to $1.47 on Hydrogen xhigh (6.5 times). Median chain wait stretched from 5.5 seconds to 37.6 seconds on Carbon, and from 2.4 seconds to 16.2 seconds on Hydrogen. Unlike our roleplay run, Argon low and medium did a sliver of thinking here—medians of 13 and 34 tokens per chain reply, against 536 on Argon high. Once Carbon and Argon list prices double on 2026-12-01 (per Google’s pricing page checked 2026-11-16), our 290-call suite climbs from $5.40 to $8.29; that page also matches the prices we used. Separately, the CLI’s output safety filter killed two rounds of our first Carbon medium chain, starting on the completely ordinary question 「你们书店最近有什么活动吗?」 (“Any events at your bookstore lately?”). Rerunning the chain in a fresh session cleared it cleanly, and the CLI gives no visibility into what tripped the filter.
How do I run a Gemini 4 companion in Foreverse?
Based on our 2026-11-16 check of the Android source code and official server model list, Gemini 4 Carbon and Argon are available on Foreverse’s official channel, while Hydrogen is not. Because AI companions run on our Agent engine, both live replies and proactive check-ins read their thinking setting from Me → Advanced settings → Agent → Agent capabilities → Reasoning effort—tap that row to cycle through Off, Low, Medium, and High (there is no xhigh). Changing Reasoning Effort inside Model providers will not affect companions, since that setting governs roleplay cards and novel continuation. Background companion tasks only force native thinking off on DeepSeek models, leaving Gemini 4 untouched; on the official channel, Low, Medium, and High map to Gemini’s thinking level while Off sends no parameter. Given our results above, set it to Low.
FAQ
Can Google Gemini 4 act as an AI companion with long-term memory?
Yes, when an app feeds persona traits and memory entries into the prompt. Across all ten Gemini 4 configurations we tested, the model answered 60 of 60 memory recall checks and respected the one painful topic the user never wanted mentioned. We did not test long conversations or large memory sets.
Which Gemini 4 model is best for an AI companion?
None pulled ahead because our test hit a ceiling across the board. The lowest-cost, fastest configurations passed right alongside the flagships: Carbon low ($0.20 per full run, 5.5 s median reply), Argon low ($0.21, 7.6 s), and Hydrogen low ($0.23, 2.4 s, shortest messages).
Does higher thinking improve Gemini 4 companion replies?
We measured no gain. High settings cost 6.5 to 7.2 times more and slowed replies roughly sevenfold without improving memory, timing, or emotional support. The lone failure—a pressuring check-in on Argon medium—did not happen on Argon low, pointing to sampling luck rather than thinking depth.
Can Gemini 4 send natural proactive first messages?
In our three timed scenarios, yes. Every one of the 30 proactive messages stayed at or below 60 characters, wove in a specific memory, and matched the time of day. Judges passed all 90 checks (four by a 2-to-1 split), though Carbon high still took a median of 59.4 seconds just to write one short ping.
Where do I change Gemini 4’s thinking setting for companions in Foreverse?
On Android, open Me → Advanced settings → Agent → Agent capabilities → Reasoning effort, and tap to cycle Off, Low, Medium, or High (xhigh is not offered). That setting controls both companion chat replies and proactive messages. Verified 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)