Flash

TopicGoogle Gemini 4

Review 5.9·Models & Products·2026-11-16 09:22

Google Gemini 4 Hydrogen Review: 2.5x Faster, Intelligence Index 51 Above Flagship Argon, Free Tier Live

Google released Gemini 4 Hydrogen on Nov 12, 2026, the third model in the Gemini 4 family, built for rapid execution in everyday coding and agentic workflows. This review checks Google's official blog and pricing page figures against independent measurements from Artificial Analysis, LMArena, and LiveBench, item by item.

A sunlit developer workspace with a laptop showing code and a phone showing search results
Gemini 4 Hydrogen review cover

Latest

1 updates · newest first
  1. Review published: Gemini 4 Hydrogen review is live.

Scorecard

5.9/ 10

HIMIAO Horizon score

A speed-tier Gemini 4 model: Intelligence Index 51 above flagship Argon, $1.50 input pricing with a free tier; terminal coding and long-agent scores still trail the in-house flagships.

Pros

  • Artificial Analysis Intelligence Index 51 (high), 54 at xhigh, above flagship Argon at 50

  • Input priced at $1.50 per million tokens, below Carbon and Argon's $2.00 (rising to $4.00 on Dec 1), and the only Gemini 4 model with a free tier

  • LiveBench 2.0 reasoning 88.6% and research 86.8%, with data analysis 83.2% ranked third

  • Google Search switched worldwide on launch day, with the Gemini app, API, and AI Studio live the same day

Cons

  • Artificial Analysis's own Terminal-Bench 4.0 run is 54.0%, below Claude Sonnet 5.5 at 64% and GPT-6.1 Sol at 56%

  • LMArena text arena 1490, ranked 15th, below Carbon at 1526 and Argon at 1522

  • LiveBench 2.0 long-agent 61.8%, more than 20 points below its reasoning and research scores

  • No SWE-bench result; Google's self-tested 2.5x speed and Terminal-Bench 4.1 score of 57% have no third-party verification yet

Best for

Developers and teams that need fast everyday coding and agent work, accept roughly 16 seconds to the first token, and value the free tier and $1.50 input pricing

01Product overview

Positioning decides the tier choice: Hydrogen is the speed model in the Gemini 4 family, with Carbon and Argon taking deep, long-horizon reasoning and Hydrogen taking everyday execution. Four thinking levels — low, medium, high (default), and xhigh — trade higher scores for more time and cost. Google has not disclosed the parameter count, and the weights are not public; the model is proprietary.

  1. Vendor and date: Google, released Nov 12, 2026; the announcement was updated Nov 13
  2. Model: Gemini 4 Hydrogen, the third model in the Gemini 4 family, a speed model for everyday coding and agentic workflows; text, image, speech, and video input, text output
  3. Thinking levels: Four levels — low, medium, high (default), and xhigh; the Artificial Analysis Intelligence Index reads 27, 41, 51, and 54 across the levels
  4. Context: 1 million input tokens and up to 1 million output tokens; flat pricing across all levels, with no long-context surcharge
  5. Pricing: Standard paid tier: $1.50 input, $7.50 output, and $0.075 cached input per million tokens; the free tier charges nothing for input, output, or cache (rate-limited); Batch and Flex at half price, $0.75/$3.75, Priority at $2.70/$13.50
  6. Availability: Gemini app, Gemini API, AI Studio, Gemini for business, Google Search (all users), Antigravity (Pro and Ultra subscribers), Android Studio, and Google Sheets (starting with Pro and Ultra)
  7. Price context: Carbon and Argon cost $2.00/$10.00 until Nov 30, 2026, rising to $4.00/$20.00 from Dec 1; 3.8 Flash costs $0.75/$3.75 until Dec 31, 2026, then matches Hydrogen at $1.50/$7.50 from Jan 1, 2027

02Measured performance

Speed is the strongest part of Google's self-tested numbers: at the default high setting, tasks finish 2.5 times faster than on Gemini 3.8 Flash; Terminal-Bench 4.1 reads 57% against 18.2% for 3.8 Flash, which Google describes as more than triple. In partners' early tests, complex graph, SQL, and data-heavy enterprise workflows ran 2x to 2.8x faster than 3.8 Flash with token use down 12% to 23%. Antigravity's Orchestrator reached near-Carbon task success rates on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster.

Agent and knowledge work is the other set of Google's self-tested numbers: HLE-Diamond at 58.8%, which Google calls leading in its class; Vals Tax Agent Bench at 52.3%, ranked first; Vals Legal Research Bench at 55.8%, tied for first according to Google; Agent Arena ranked fifth; Agents' Last Exam v1.1 at 71.6%; the multimodal knowledge agent benchmark MKA at 53.4%; and fabricated claims down more than 5x against 3.8 Flash.

Independent leaderboards read separately from vendor numbers: the Artificial Analysis Intelligence Index v5.0.1 gives 51 (high) and 54 (xhigh), above Argon's 50; on the evaluator's own Terminal-Bench 4.0 run, Hydrogen scores 54.0% (high), below Claude Sonnet 5.5 at 64% and GPT-6.1 Sol at 56%. LMArena's text arena (style-controlled) gives 1490, ranked 15th, below Carbon at 1526 and Argon at 1522, with WebDev HTML at 1743, ranked fifth. LiveBench 2.0 gives 77.0 overall, ranked ninth, with reasoning 88.6%, research 86.8%, coding 78.5%, data analysis 83.2% (ranked third), and long-agent work 61.8%, at $1.924 per successful task. Artificial Analysis measured 8.3 minutes (high) and 14.7 minutes (xhigh) per Intelligence Index task, $1.87 and $3.16 per task, and 16.21 and 42.26 seconds to the first token. CursorBench 4.0 gives 54.3% (second-hand report), above 3.8 Flash at 39.6%.

The following readings must stay separate: no SWE-bench version has measured Hydrogen; Google cites its own SWE-sweep variant at 13.5%, not the public SWE-bench, so this review carries no SWE-bench score. The two Terminal-Bench versions do not compare directly: Google's self-run 4.1 reads 57%, while independent evaluators measured 54.0% on 4.0. The 55.8% for Vals Legal Research Bench is Google's cited value; another second-hand source puts the leaderboard at 53.9% at the time, and the figures are not reconciled. Artificial Analysis time per Intelligence Index task at xhigh follows the model page's current value of 14.7 minutes; the 13 to 13.5 minutes in early write-ups is an older value.

03Scorecard

The conversion method is stated openly: take five representative scores from Google's published figures and divide by 10 for a 10-point scale; the total is the mean of the five converted scores. Converted scores restate benchmark figures only, with no editorial judgment added; independent leaderboard scores are listed separately and do not enter the total.

  1. Terminal coding 5.7: Terminal-Bench 4.1 score of 57 converted; source: Google's official blog figures
  2. Hard knowledge 5.9: HLE-Diamond 58.8 converted; source: Google's official blog figures
  3. Professional workflows 7.2: Agents' Last Exam v1.1 score of 71.6 converted; source: Google's official blog figures
  4. Tax agent 5.2: Vals Tax Agent Bench 52.3 converted; source: Google's official blog figures
  5. Multimodal knowledge 5.3: MKA 53.4 converted; source: Google's official blog figures
  6. Total 5.9: Mean of the five converted dimension scores (5.7 + 5.9 + 7.2 + 5.2 + 5.3) ÷ 5
  7. Independent leaderboards, listed separately: Artificial Analysis Intelligence Index v5.0.1 at 51 (high) and 54 (xhigh), above Argon at 50; Artificial Analysis's own Terminal-Bench 4.0 run at 54.0% (high); LMArena text arena 1490 and WebDev HTML 1743; LiveBench 2.0 overall 77.0, with reasoning 88.6%, research 86.8%, coding 78.5%, and long-agent work 61.8%; none of the four enters the total

04Who it fits

On features and price, the following users can already reach the model through public channels: developers who need fast everyday coding, repository debugging, and multi-step agent tasks, with the free tier and API both live and standard paid input at $1.50 per million tokens; teams handling long documents, codebases, and multimodal files, with 1 million input tokens and up to 1 million output tokens live; and everyday Google Search users, who need to do nothing — Search has switched worldwide.

In the other direction, long-horizon deep reasoning stays with Carbon and Argon, while Hydrogen is positioned for rapid execution. For interactive settings sensitive to first-token latency, Artificial Analysis measured 16.21 seconds at high and 42.26 seconds at xhigh, which belongs in any selection decision. With no SWE-bench result yet, repository-scale coding choices need re-testing on the target task first.

Why it matters

Hydrogen is the only Gemini 4 model with a free tier, scores 51 on the Intelligence Index above flagship Argon, and now powers Google Search worldwide.
Useful Tap if this story helped you
Send to a friendMakes a share image for X, WeChat, Telegram and more

SourcesGoogle official blog (Nov 12, 2026); Google Search announcement (Nov 12, 2026); Google official pricing page (checked Nov 16, 2026); Artificial Analysis model page (checked Nov 16, 2026); LMArena text leaderboard (Nov 11, 2026); LiveBench 2.0 leaderboard (checked Nov 16, 2026); TechCrunch (Nov 12, 2026). This article is compiled from public information and does not constitute investment advice.

From HIMIAO Horizon · himiao.ai

Grab todayOpenAI Codex for Students: $100 in Free CreditsHIMIAO Picks →

Follow HIMIAO Horizon

No sign-up · Unsubscribe any time

Email

Notifications

A ping on your phone or computer when the brief is out. No email needed.

HIMIAO PicksWeekly freebies: one email every Friday, a 5-minute read

Comments

  1. Loading comments…
Ask the cat