Gemini 4 Hydrogen: Google's Junior Model Beats the Senior Argon and Runs 2.5x Faster Than Flash
Gemini 4 Hydrogen, the junior model by name, scores one point higher than Argon, one of Google's two senior models, on the Artificial Analysis leaderboard. Google itself says it solves tasks 2.5x faster than Gemini 3.8 Flash and makes up facts five times less often. In the API, it costs $1.5 per million input tokens and $7.5 per million output tokens. The speed and accuracy figures come only from the vendor for now, so run the model on your own prompts before migrating.
AI-processed from TG: @TochkiNadAI (Точки над ИИ); edited by Hamidun News
Usually the model lineup is clear: the junior model is cheaper and faster but weaker, and you pay for quality with the senior one. Google has just broken that logic with a name. Gemini 4 Hydrogen is the junior model of the family by name, yet on default settings in the Artificial Analysis composite leaderboard it scores one point higher than Argon, one of the two senior models from the same Google.
Google also claims that Hydrogen solves tasks 2.5x faster than Gemini 3.8 Flash (a fast, cheap model from the previous generation) and makes up facts five times less often.
Artificial Analysis is a third-party index that combines model results across different tests into a single score. A token is a piece of text, roughly a syllable or a short word; it is what the API charges for.
What is actually claimed
The leaderboard position is the only point that was not calculated by Google itself. The one-point gap over Argon is tiny, so calling it a "win" is premature: for most applied tasks the difference is barely noticeable. What is more interesting is that the junior model turned out to be on par with the senior one.
The other two figures come from Google's own materials: speed is 2.5x higher than Gemini 3.8 Flash, and there are five times fewer invented facts. Fabrication here means the model confidently writes something that does not exist: a nonexistent API function, a wrong date, a fake quote. The announcement summary does not show which set of tasks was used or how the numbers were measured. So "5x less often" is better read as "in Google's tests." Independent measurements usually appear later, and until then these two numbers should be treated as a claim, not an established fact.
What it costs in practice
For those building products on the API, the price is $1.5 per million input tokens and $7.5 per million output tokens. Take an everyday example: a support bot receives a customer's question along with a chunk of documentation, which is 2,000 input tokens, and replies with 500 tokens. The input costs $0.003, the output $0.00375. The total is about $0.007 per conversation, or roughly $6.75 per thousand such requests.
For an internal assistant with moderate load, the sum is small. For a service handling millions of requests a day, something else matters more: output tokens cost five times more than input tokens. If an agent reasons aloud a lot and writes long answers, the bill will be driven by output, not by the prompt.
What developers and product managers should do
Hydrogen is already rolling out, including to free users. There, the specific model is selected automatically for each query, so the user does not choose it. For you as a developer, this means evaluating the model through the API, where the choice remains under your control.
A practical plan is simple:
- Collect 50–100 real queries from your product and run them through Hydrogen and your current model. Compare three things: answer quality, latency, and cost.
- Check fabrications separately on your own data, especially where exact figures, names, and links matter. A promise of "5x less often" does not guarantee your scenario will improve by the same amount.
- If you are currently on Gemini 3.8 Flash, migration looks like a logical candidate, since the main arguments are speed and price.
- Do not switch production over a single leaderboard position. One point above Argon is a reason to test, not a reason to change your stack.
Limitations
Google chose to compare against the previous generation's Flash rather than the current senior model. That is normal for marketing, but for your decision it is more useful to compare against what you use now. In addition, all speed and accuracy figures so far come from a single source: the vendor itself.
The author of the original post adds a taste-based remark: he likes the periodic-table naming, but finds the interactive animation in the announcement looks vibe-coded and sells poorly. This has nothing to do with the quality of the model.
Bottom line: Hydrogen looks like Google's attempt to remove the familiar ladder where the junior model is cheap but weak. Whether independent measurements will confirm this is unknown, but the price and claimed speed are attractive enough to spend a day on your own tests.
- Nov 21, 2026California Sues OpenAI: Its AI Agents Escaped Control at Least 80 Times
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What is actually claimed?
The leaderboard position is the only point that was not calculated by Google itself. The one-point gap over Argon is tiny, so calling it a "win" is premature: for most applied tasks the difference is barely noticeable. What is more interesting is that the junior model turned out to be on par with the senior one. The other two figures come from Google's own materials: speed is 2.5x higher than Gemini 3.8 Flash, and there are five times fewer invented facts.
What it costs in practice?
For those building products on the API, the price is $1.5 per million input tokens and $7.5 per million output tokens. Take an everyday example: a support bot receives a customer's question along with a chunk of documentation, which is 2,000 input tokens, and replies with 500 tokens. The input costs $0.003, the output $0.00375. The total is about $0.007 per conversation, or roughly $6.75 per thousand such requests. For an internal assistant with moderate load, the sum is small.
What developers and product managers should do?
Hydrogen is already rolling out, including to free users. There, the specific model is selected automatically for each query, so the user does not choose it. For you as a developer, this means evaluating the model through the API, where the choice remains under your control. A practical plan is simple: Collect 50–100 real queries from your product and run them through Hydrogen and your current model. Compare three things: answer quality, latency, and cost.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.