Independent unofficial site. Not affiliated with Google, Google DeepMind, or Alphabet. Gemini is a trademark of Google.
The Hydrogen Brief An independent, unofficial guide EN · 中文

We also ran our own tests. Original tests / 我们的原创实测 → The cards below are community posts, not re-run by us.

Same prompt, models side by side

18 public runs where people sent one prompt (or one task) to Gemini 4 Hydrogen and other models and posted the results. 6 include dollar costs. Every number is the author’s own claim, copied from the linked post on 14 Nov 2026. We did not re-run them, settings differ between posts, and faster or cheaper is not the same as better: several authors preferred the slower model’s output.

Run your own test free in AI Studio ↗Prompts you can copyFree access guide

Three.js Spider-Man model — API cost and working time — result card

Three.js Spider-Man model — API cost and working time

Author says: same prompt · All at highest effort (author) · @VibeDevarf, 2026-11-14

ModelTimeCost / quota
GPT 6 Astra1h 25m$7.37
Claude Opus 5.52h 13m$38.45
Claude Fable 5.22h 03m$41.12
Gemini 4 Argon3h 51m$8.13
Gemini 4 Carbon1h 20m$12.72
Gemini 4 Hydrogen1h 06m$18.08

Same prompt for all six, per the author; building is part of the prompt. Hydrogen was the fastest here but not the cheapest. The author's own run, not a benchmark.

Download card
UP house — same reference, one shot — result card

UP house — same reference, one shot

Author says: same prompt · XHigh (author label) · @noclipepe, 2026-11-13

ModelTimeCost / quota
Gemini 4 Hydrogen12m 49s~$1.09
Gemini 4 Carbon24m 45s~$1.90
Claude Opus 5.536m 46s~$4.20
Claude Fable 5.245m 44s$19.10

Same prompt, same reference, XHigh, one shot — per the author. Fable made the author's best house; Hydrogen the 'haunted version'. Low cost is not the same as best quality.

Download card
Night Tokyo — one prompt, four labs — result card

Night Tokyo — one prompt, four labs

Author says: same prompt · Mixed: Fable medium; Astra / Hydrogen xhigh; Orchestrator · @sprinntt, 2026-11-13

ModelTimeCost / quota
Fable 5.2 (medium)20 min$7.11
GPT-6 Astra (xhigh)13 min$2.11
Hydrogen (xhigh)20 min$2.25
Orchestrator (Argon+H)14 min$1.29

The author says one prompt, but reasoning levels and tools differ. Astra looked plainer yet had the most detail up close, per the author. Not a controlled benchmark.

Download card
Business blog article (2,000–2,500 chars) — blind AI scoring — result card

Business blog article (2,000–2,500 chars) — blind AI scoring

Author says: same prompt · Hydrogen / 3.8 Flash: MEDIUM; Carbon: LOW (author labels) · @domodomo, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen24.9 s¥4.9
Gemini 3.8 Flash24.4 s¥1.2 (list ¥2.4)
Gemini 4 Carbon30.6 s¥2.5 (list ¥5.0)

Same system prompt and task for all models, per the author; outputs were anonymised and scored twice by an AI judge (Hydrogen 98, 3.8 Flash 97.5, Carbon 83). Hydrogen's 3,143 thinking tokens drove its higher cost. The author's runs, not a benchmark.

Download card
800-character summary of a subsidy document — blind AI scoring — result card

800-character summary of a subsidy document — blind AI scoring

Author says: same prompt · Hydrogen / 3.8 Flash: MEDIUM; Argon: LOW (author labels) · @domodomo, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen21.0 s¥4.7
Gemini 4 Argon18.6 s¥1.6 (list ¥3.3)
Gemini 3.8 Flash23.5 s¥1.6 (list ¥3.2)

Same instruction for all models, quoted in the post ('only from the material… write 資料からは不明 where unclear'). Scores per the author's AI judge: Hydrogen 97.5, Argon 95.5, 3.8 Flash 84.5 (it omitted the required 'unknown' note). Not a controlled benchmark.

Download card
3D Star Core Temple challenge — time to finish — result card

3D Star Core Temple challenge — time to finish

Author says: same prompt · @BadEarsX, 2026-11-14

ModelTimeCost / quota
DeepSeek V4.1 101750 minutesnot stated
Qwen4 Max Preview56 minutesnot stated
Gemini 4 Hydrogen3 minutes 30 secondsnot stated

Same prompt, per the author, who said Hydrogen's visuals were less polished than DeepSeek's but it was the most stable (no black screens or flicker). The author's own run, not a benchmark.

Download card
Hydrogen vs Carbon — timed and billed — result card

Hydrogen vs Carbon — timed and billed

Same task (prompt not shown identical) · High · OpenRouter · @EnvolDev, 2026-11-12

ModelTimeCost / quota
Gemini 4 Hydrogen1h 23m 40s$4.27
Gemini 4 Carbon2h 39m 10s$8.83

Author's OpenRouter run at High. The post reports time and price but does not publish the prompt; treat as a same-task demo, not a controlled same-prompt test.

Download card
Hydrogen vs DeepSeek v4.1 Flash — same prompt — result card

Hydrogen vs DeepSeek v4.1 Flash — same prompt

Author says: same prompt · @rainwaterJay, 2026-11-12

ModelTimeCost / quota
Gemini 4 Hydrogen3 min$1.30
DeepSeek v4.1 Flash 1017100 min$0.90

Same prompt, per the author, who set quality aside: Hydrogen was much faster but cost more in API credits in this run.

Download card
Snow village — result card

Snow village

Author says: same prompt · High · @badboyfoxy, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen~3 min$1.37
GPT-6.1 Sol~15 min$4.03

The author says Hydrogen was faster and cheaper but did horribly on the scene; Sol was more creative. Harness and surface are not controlled. These are the author's runs, not a benchmark.

Download card
Hydrogen vs GPT-6.2 Sol — quota used — result card

Hydrogen vs GPT-6.2 Sol — quota used

Author says: same prompt · @Stargridcompass, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen3 min2% of 5-h quota
GPT-6.2 Sol13 min10% of 5-h quota

Same prompt set, per the author: Sol handled detail better; Hydrogen was 4x+ faster and used a fifth of the quota. Quota is the author's subscription window, not dollars.

Download card
3D star-core temple — result card

3D star-core temple

Author says: same prompt · @YxzRainy, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen~3m 30s+not stated
DeepSeek V4.1 101750 minnot stated
Qwen4 Max Preview56 minnot stated

The same test, according to the author. Hydrogen had less texture/detail but was most stable in this run; cost and thinking settings are not stated.

Download card
Clickable black liquid with golden ripples — result card

Clickable black liquid with golden ripples

Same task (prompt not shown identical) · @Q1n9H, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen1 minnot stated
Gemini 3.8 Flash5 minnot stated
GPT-6.1 Sol7 minnot stated

Author reports these timings for the same liquid-surface task; Hydrogen's scene was not as good as Sol's. No cost, exact prompt or effort setting published.

Download card
LEGO world build — result card

LEGO world build

Same task (prompt not shown identical) · @kepochnik, 2026-11-13

ModelTimeCost / quota
Gemini 4 Hydrogen28 minnot stated
Claude Fable 5.244 minnot stated

Author-reported build times for the same LEGO-world task; the author asks viewers which did better. No cost stated.

Download card
Hydrogen vs Fable 5.2 — weekly quota — result card

Hydrogen vs Fable 5.2 — weekly quota

Author says: same prompt · @Crypto_146827, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogennot stated1% of weekly quota
Claude Fable 5.2not stated15% of weekly quota

Same prompt set, per the author, who judged Fable more expressive and said Hydrogen was 3x+ faster. Exact times not stated.

Download card
Ship in a bottle — result card

Ship in a bottle

Author says: same prompt · @Betaer, 2026-11-14

ModelTimeCost / quota
Gemini 4 Hydrogen4m 47snot stated
Gemini 4 Carbonnot statednot stated

Same prompt, says the author, who preferred Hydrogen's result. Only Hydrogen's time is stated; Carbon's time and both costs are unknown.

Download card
Oil painting of a rocket — result card

Oil painting of a rocket

Same task (prompt not shown identical) · High · @gwammyy, 2026-11-13

ModelTimeCost / quota
Claude Sonnet 5.5not statednot stated
Gemini 4 Hydrogennot statednot stated

The author posted left/right outputs labelled Sonnet 5.5 (high) and Hydrogen (high), with the short request 'Oil painting of rocket'. No time or cost is stated.

Download card
Pelican riding a bicycle — interactive 3D — result card

Pelican riding a bicycle — interactive 3D

Same task (prompt not shown identical) · Flash / Argon / Carbon: Max; Hydrogen: Extra (author labels) · @StatsWire, 2026-11-14

ModelTimeCost / quota
Gemini 3.8 Flashnot statednot stated
Gemini 4 Argonnot statednot stated
Gemini 4 Carbonnot statednot stated
Gemini 4 Hydrogennot statednot stated

Four outputs of the pelican task, not four measured speeds. The author did not publish time, cost or full prompt text. Effort labels differ. Watch the embed; no winner is inferred.

Download card
Three models, one prompt — 3D globe — result card

Three models, one prompt — 3D globe

Author says: same prompt · @jayakavuru, 2026-11-14

ModelTimeCost / quota
Fable 5.2not statednot stated
GPT-6.2 Solnot statednot stated
Gemini 4 Hydrogennot statednot stated

The author explicitly says 'same prompt', but does not share the prompt or times/costs. They say all three have strengths. Model names are kept as the author wrote them, including GPT-6.2 Sol.

Download card

Try these prompts (shared publicly by their authors)

Only prompts the author posted themselves, copied verbatim with credit. “Try free” copies the prompt and opens Google AI Studio with Hydrogen selected; paste it in. AI Studio’s free tier is rate-limited and free-tier prompts are used to improve Google products.

Prompt by Simon Willison (@simonw) · simonwillison.net

Generate an SVG of a pelican riding a bicycle

Simon Willison's public LLM benchmark prompt. Many Hydrogen pelican posts (e.g. @StatsWire, @hiwise_aigc) build on it.

Prompt by 产品晓东 (@qingzhoucloud) · X

创建一个 HTML,内容是 SVG 绘制一个鹈鹕骑自行车的 2D 动画,你不需要任何测试,不要有任何限制,【尽你最大的能力】

Shared by the author with a warning: adding the final phrase 【尽你最大的能力】 ("do your very best") made their result worse. Try it with and without that phrase.

Prompt by V哥来了 (@vincent_ddup) · X

创建一款 2D 平台跳跃游戏,包含奔跑、跳跃、二段跳、移动平台、金币和危险障碍。加入检查点系统、金币计数器,以及抵达后即可结束关卡的终点旗帜。使用方向键或 WASD 实现灵敏顺畅的操作,并加入简单的关卡难度选择功能,或在角色死亡后提供重新开始选项。

The author ran this 2D platformer prompt (which they say is the Grok Build example prompt) in the Gemini app with Hydrogen: 6 minutes and 3 revisions by their count.

Prompt by DANJI (@DanjiTosaka) · X

Create a single-file, self-contained Three.js (ES module via importmap) AAA-style 3D interactive web experience titled "天狗 TENGU'S SANCTUARY" in a single `index.html` with zero external assets (all textures, geometry, shaders, and audio generated procedurally).

1. WORLD, TERRAIN & ARCHITECTURE
- Build a dramatic Zhangjiajie-style karst valley & canyon basin with layered granite/sandstone mountain spires, natural stone arches at the canyon entrance, terraced cliff plateaus, waterfalls with animated plunge-pool splash rings/mist, and a solid dark-teal lakebed floor beneath the water so no empty void is ever visible.
- Populate the valley with traditional East-Asian architecture merged via a custom GeometryBatcher: a central multi-tiered pagoda, a Stone Arch Bridge, a grand Torii gate, a Fortress Gate, Lantern Town stilt/machiya/inn houses, harbor piers, stone stairways, stone lanterns, and a rotating watermill.
- Add lush procedural vegetation: gnarled Sakura cherry blossom trees, layered niwaki pines (including a guardian pine at the summit sanctuary), autumn maples, bamboo groves, lotus/lily pads on the water, and ~6,500 instanced wind-animated grass tufts.

2. HERO CHARACTER: GREAT FOREST DAI-TENGU
- Perched solidly on a stepped granite sanctuary plinth with a mossy boulder spur and twin Shimenawa votive pillars at the top central peak (no floating/bouncing).
- Sculpted via SDF marching cubes + Lathe/TubeGeometry: pleated hakama trousers, white kyahan leg wraps, single-tooth Tengu-geta sandals, athletic torso framed by a crossed-lapel forest-green Yamabushi robe with crimson/gold brocade trim, braided Shimenawa rope sash with paper shide charms, Yuigesa pom-poms, and Juzu prayer beads.
- Sculpted SDF Dai-Tengu face with long aquiline nose, dark eye sockets with glowing amber slit hawk eyes, feathered silver-shaded eyebrows, 22 wind-swept silver hair locks, hexagonal black-and-gold Tokin hat, clawed hands holding a 9-lobe jade-and-gold Ha-Uchiwa fan and a gold-ringed Shakujo pilgrim staff, plus 4-tier raven-black feathered wings oriented with quaternions.

3. LIVING VILLAGE, KOI & INTRO BOAT RIDE
- Build visible, grounded stone/wooden pathways and supported stairs connecting the Stone Bridge, harbor piers, Lantern Town, Fortress Gate, and East Falls, plus a unified `getWalkableHeight(x, z)` surface sampler (matching the bridge's parabolic arch).
- Add ~30 living villagers (kimono/yukata, straw kasa hats, monks, lantern-bearers, merchants) walking along those exact paths with smooth walking bob and limb swing, plus stationary monks/guards.
- Add 18 animated Koi fish swimming beneath the clear jade water, flying white cranes, floating sky/water lanterns, and drifting sakura petals/golden forest pollen.
- Add a Chinese wooden sampan boat (`introBoat`) with bamboo canopy, bow lantern, boatman, and V-shaped water wake that glides from outside the canyon arch (`z = 168`) into the village (`z = 74`).

4. CINEMATOGRAPHY, FIRST-PERSON MODE & ANTI-CLIPPING
- Start on a 24-second 3-act cinematic camera sequence: wide high zoom-out framing both the entering sampan and the Dai-Tengu on the summit, gliding down to the Stone Bridge, then rising to a close-up of the Dai-Tengu. Detach immediately on any user mouse drag, wheel, preset click, or First-Person toggle.
- Enforce camera anti-void and anti-rock-clipping clamping every frame (never below water/terrain, never inside hollow cliff plateaus).
- Include a playable "First-Person Walk (WASD)" mode with PointerLock/drag mouse look, Space jump, Shift sprint, and `getWalkableHeight` ground following.

5. AAA RENDERING, MOODS, AUDIO & 60FPS VIDEO EXPORT
- Custom GLSL sky dome with azure daylight gradient and horizon cumulus clouds, planar water reflections, soft PCF shadows, half-res bilateral SSAO, world-space volumetric light shafts, UnrealBloomPass, and ACES Filmic split-tone post-processing.
- Three instant mood presets: "Jade Valley & Sun" (default crisp high-contrast daylight), "Twilight & Lanterns", and "Tengu Eclipse".
- Minimalist English HUD with collapsed-by-default "AAA ENGINE PARAMETERS" drawer, camera preset bar, and a "🎬 Record Video" button using `canvas.captureStream(60)` + `tenguAudio.getAudioStream()` via `MediaRecorder` (`vp9,opus`) to export a 60fps `.webm` video with audio.
- Procedural Web Audio engine (`tenguAudio`) unlocked on user interaction: Karplus-Strong physical-modeled Shamisen/Koto, noise-burst + sub-bass Epic Taiko war drums, breathy Shakuhachi flute, stereo pink-noise pine wind & waterfall ambience dynamically modulated by camera altitude and waterfall proximity.

Long Three.js prompt the author shared publicly for their Tengu's Sanctuary scene (made with Hydrogen in swarm mode). Single-pass chat in AI Studio may differ from their multi-agent run.

Prompt by 虚惊一场|Builder (@R7Ruk) · X

根据这个视频,画出这个房型图纸

One-line prompt the author used with a first-person walkthrough video of a home; attach your own video when trying it.

Official speed and price numbers

For measured tokens per second and cost per task from Artificial Analysis, and official list prices, see the comparison hub and the cost calculator.