All News
googlegemini-4-hydrogenbenchmarkspricingagentic-workflowsapi

Gemini 4 Hydrogen leads its sibling by a point. The gap that matters is elsewhere

Google's Gemini 4 Hydrogen beats Argon by one point on the index its launch post cites. Independent measurements put the real gap in time and cost per task.

Vlad MakarovVlad Makarovreviewed and published
6 min read
Gemini 4 Hydrogen leads its sibling by a point. The gap that matters is elsewhere

Google introduced Gemini 4 Hydrogen on Thursday, and the launch post by Koray Kavukcuoglu opens with a comparison that reads like a filing: Hydrogen "surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level (51 vs. 50, and reaching 54 at its highest xhigh thinking setting)." One point over Google's own previous model, plus a fourth that appears only at deeper reasoning, is not a capability leap. It is a claim about where Hydrogen sits in the lineup, and the figures that would describe that position — output speed, cost per task — appear nowhere in the post.

The index it wins and the sibling that stays ahead

Hydrogen's 51 is real. Artificial Analysis lists the high-effort model at 51 on Intelligence Index v5.0.1, twenty-first of eighty tracked, with Argon at the same effort on 50. The sentence omits that the frontier model in the family does not lose: Carbon reads 60, nine points clear of Hydrogen's default and six above its xhigh ceiling. Hydrogen beats the sibling it also undercuts on price, not the one Google calls strongest. The same pages describe an efficiency picture that is considerably less ambiguous than the index.

Artificial Analysis, index v5.0.1IndexTime per taskCost per task
Gemini 4 Carbon (high)6028.6 min$5.48
Gemini 4 Hydrogen (xhigh)5414.1 min$3.16
Gemini 4 Hydrogen (high, default)517.8 min$1.87
Gemini 4 Argon (high)5030.1 min$4.97

The speed number Google left to somebody else

The post promises Hydrogen completes tasks "faster than models in its class" and "2.5 times faster" than Gemini 3.8 Flash. Neither claim carries a unit. There is no tokens-per-second figure, no time-to-first-token figure and no absolute latency anywhere in fifteen minutes of reading, which makes the independent measurement worth more than the marketing: 181 output tokens per second against 64 for Argon, 16.3 seconds to first answer token against 42.7, and 7.8 minutes per index task against 30.1. A fourfold wall-clock difference, same harness, same index.

Google led with the one number where it is indistinguishable from its sibling and buried the axis where it is not. The xhigh tier also carries a cost the post states only as quality: one extra index point moves time to first token from 16.3 to 42.9 seconds, index output from 110 million tokens to 304 million, and cost per task from $1.87 to $3.16. Artificial Analysis ranks that setting seventy-eighth of seventy-nine on verbosity.

What the scorecard shows, and what it only animates

Google's benchmark comparison ships as an animated GIF, so the per-model bars cannot be read as text. The prose figures are a mixed set: 57% on Terminal-Bench 4.1 against 18.2% for 3.8 Flash, 58.8% on HLE-Diamond, 71.6% on Agents' Last Exam v1.1, 53.4% on the Multimodal Knowledge Agent benchmark, first place on Vals Tax Agent Bench at 52.3%, tied first on Vals Legal Research at 55.8%, fifth on Agent Arena.

Two of those need labels. MKA is Google's own benchmark, introduced the same week. AutomationBench-AA is claimed as a class lead with no number printed at all, and fifth place in an arena is a snapshot of a live board. Customer results arrive anonymized in the summary — "our partners have seen Hydrogen run 2x to 2.8x faster while reducing token usage by 12% to 23%" — though the quotes beneath name Box, STAN and Pixelcut. No methodology, nothing reproducible from outside.

Flat pricing is the actual decision

The figure with no interpretation problem is the price: $1.50 per million input tokens and $7.50 per million output, flat across the full 1-million-token window. Carbon and Argon both sit at $2 and $10, per Google's Carbon post, so Hydrogen is 25% cheaper on both sides of the meter. The flat structure matters as much as the level. A request at 900,000 tokens costs the same per token as one at 9,000, which is what makes the window usable rather than decorative, and Artificial Analysis records a 95% cache discount on top.

Orchestrator, described in a separate Antigravity post and covered at Carbon's launch, hands control between Carbon and Hydrogen mid-session. Google reports near-Carbon success on Terminal-Bench with 24% less quota consumed and 19% faster completion. The two posts disagree about which version of that benchmark was run: the main announcement says 4.0, the Antigravity post says 4.1.

Everyone, with an asterisk attached

Hydrogen reaches all tiers of AI Studio, the Gemini API, Gemini for business, the Gemini app, Search, and paid Google AI Pro and Ultra subscribers in Antigravity. Google's word is "everyone." A footnote narrows it: free users in the app get Hydrogen only when auto model selection decides it is the right model for the prompt, which is a router choosing, not a dropdown. There are no weights, no licence terms and no repository — the only technical artifact is a model card PDF. Hydrogen is a rented endpoint with a price list, and the post does not pretend otherwise.

Safety claims that stay inside the building

The safety section is dense with precise-looking percentages: up to 81% less overreach in agentic tasks, 79% less under-service, a 12x cut in misaligned actions and task shortcutting, a 94% honesty rate, 99.9% chain-of-thought legibility across "more than 200,000 audited turns," 2.3x resilience on the Gray Swan prompt-injection benchmark, false refusals under 0.4% on HealthBench Pro.

All of it is internal. None has been replicated outside, and the 200,000-turn audit was run by the party reporting it. For a model pitched at autonomous execution, writing code and touching repositories, those are the numbers that would matter most if they could be checked.

What would change the read is a placement on the coding-agent index that has favoured Anthropic, and independent runs beyond a general composite, the way the Gemini Decision 1 review had to weigh a vendor scorecard against two losses on tool use. For now the buying case is narrow: teams leaving Gemini 3.8 Flash get a large jump in completion and speed at a lower token price, and teams already buying at Carbon-tier prices get a worker at $1.87 per index task against Argon's $4.97, 62% less. Neither argument requires believing that 51 is meaningfully better than 50.

Related Articles

Scroll down

to load the next article