This check sends no prompt to Gemini 4 Hydrogen. It rests on Google's announcement, its model card and the Gemini API model page, outside boards read between 05:53 and 06:12 UTC on 13 November 2026, and 106 first-hand user reports from about the first 12 hours after launch. The result: seven of the eleven claims match the owners' boards, two match after rounding, and two do not; Google's "#5 on Agent Arena" and the full Safety & Security Report could not be checked.
On Artificial Analysis's Intelligence Index v5.0.1, read 13 November 2026, Gemini 4 Hydrogen scores 50.787 at high against Gemini 4 Argon's 50.412 — Google's announcement figure of "51 vs. 50" is the rounded display; xhigh at 53.919 is the announcement's 54. Cost per index task is $1.87 against $4.97, about 0.38 times, and 7.90 minutes per task against 29.96. Note Argon's $4.97 is at its introductory $2 / $10, which doubles on 1 December, widening the gap. On the same table, Carbon at high scores 60.026 for $5.48 per index task — 9.239 points above Hydrogen for about 2.9 times the cost. The model card adds a sentence worth keeping alongside: its MCC safety assessment found Hydrogen showed no material capability increase over Argon — a safety reading, while the index is a benchmark average; both texts stand.
Google says Hydrogen "ties for #1 on the Vals Legal Research Bench (55.8%)", but the Vals board, updated 11 November and read 13 November, lists Hydrogen at 53.85%, sixth of 95 — the 55.77% at the top is Gemini 4 Carbon's. The 53.4% on MKA is Google's own run: Artificial Analysis's MKA board has no Hydrogen value, and Claude Fable 5.2 (56.56) and Carbon (54.34) both sit above 53.4. The other seven claims match: Vals Tax 52.3% at #1, HLE-Diamond 58.8%, Vending-Bench 2 at $14,549, AutomationBench-AA 76.14%, and 3.8 Flash's Terminal-Bench 4.1 at 18.2%, among others. Hydrogen's own 57% on Terminal-Bench 4.1 is likewise Google's own run — Harbor's official board has no Gemini 4 row. One independent re-check is also worth recording: LLM Leaderboard's xhigh reruns come in below Google's figures — MKA 49.52% against Google's 53.4%, Terminal-Bench-Science 49.52% against 51.4%.
"2.5 times faster than Gemini 3.8 Flash" only holds on Artificial Analysis's launch-evening data: a Wayback copy from 21:26 UTC on 12 November shows Hydrogen at high taking 6.01 minutes per index task against 14.83 for 3.8 Flash at high, 2.47 times faster — while the live figures on 13 November give 7.90 against 14.71, 1.86 times. Output speed is 179.2 tokens per second against 112.1 for 3.8 Flash, 1.60 times. Google's partner figure, "2x to 2.8x faster", has no published source. One more independent speed reading: LLM Leaderboard puts Hydrogen at 198.2 tokens per second, against 144.0 for Sonnet 5.5 and 136.6 for 3.8 Flash.
AA index and cost per task by level: low 26.894 / $0.29, medium 41.233 / $0.80, high 50.787 / $1.87, xhigh 53.919 / $3.16. Xhigh adds 3.13 index points over high for $1.29 more per index task, 1.73 times the minutes and 2.75 times the output tokens, 303.6M against 110.4M; Artificial Analysis calls the setting "slower than average per task and very verbose", 76th of 77 on verbosity. A user report agrees: on a real production workload, xhigh used about three times the thinking tokens of high. Google's guide calls high "Balanced intelligence and speed" and says to reserve xhigh for hard, multi-step coding and agentic problems where high falls short.
$1.50 / $7.50 per million tokens, flat across the 1,048,576-token window; Batch and Flex cost $0.75 / $3.75; cached input costs $0.075 through implicit caching, which runs by default on prompts of at least 4,096 tokens, and explicit caching is not supported; thinking tokens are billed as output. From 1 December, Argon and Carbon rise to $4 / $20, so Hydrogen's list price falls from 0.75 times theirs to 0.375 times. Gemini 3.8 Flash stays at $0.75 / $3.75 through 31 December 2026 and costs $1.50 / $7.50 from 1 January 2027 — the same as Hydrogen. The free tier is free of charge and rate-limited on the Gemini API and in Google AI Studio, with grounding with Google Search not available; free Gemini app users get Hydrogen only when Google's auto selection picks it.
106 first-hand day-one reports — 70 on Reddit, 24 on Hacker News and 12 on Google's AI developer forum — mostly praise speed and light subscription quota, split on quality against Argon and Claude Sonnet 5.5, and complain about Antigravity suspensions. One user measured 37% more expensive per task than 3.8 Flash but 40–60% of the total time per task, both models on medium reasoning. After the 12 November Antigravity update, a developer-forum topic collected 40 posts between 19:31 and 22:03 UTC from users whose accounts returned "This service has been disabled in this account for violation of Terms of Service", several right after selecting Hydrogen; several posters reported access back between 21:02 and 21:43 UTC. A forum account signing "Google Support" called it a "known issue", but the account has no staff flag and no Google staff account confirmed a cause in anything we read, so we state none.
Knowledge cutoff March 2026; the model card says Hydrogen is "adapted from Gemini 4 Carbon". Unjustified refusals rose 5.57 percentage points against Gemini 3.8 Flash on Google's automated test, particularly for multimodal prompts. Under its Frontier Safety Framework, Google ran a Material Capability Change assessment and Hydrogen showed no meaningful new capabilities over Argon, so a full suite of critical capability evaluations was not required; targeted tests found it does not reach the CBRN or cyber critical capability levels, though it reaches the alert thresholds for both. The full Safety & Security Report was not out on 13 November 2026.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Medium — Gemini 4 Hydrogen review: does it beat Gemini 4 Argon, what does it cost, and which of Google's claims hold? (Nov 13, 2026)
← Back to the Gemini 4 Hydrogen topic page