Gemini 4 Hydrogen: Google's top level from September for a third of the cost. What it means for your AI bill

A man with grey hair and a beard in a black knit sweater lets a small, bright orb of light float above his open hand. Next to him a much larger orb of the same brightness rests heavily in a cushion of crumpled tracing paper. Next to him a card: 62 percent lower cost per task than Google's Gemini 4 Argon, at the same level. Gemini 4 Hydrogen from Google, Nov 12, 2026. Intelligence Index at the default level 51 vs. 50, time per task 7.9 vs. 30 minutes, cost per task $1.87 vs. $4.97. Measured by Artificial Analysis. OpenAI's GPT-6.1 Sol does the same for $0.59

TL;DR. On November 12, Google released Gemini 4 Hydrogen, a fast model for everyday work with agents and code. It costs $1.50 per million input tokens and $7.50 per million output tokens. In the independent measurement by Artificial Analysis it scores 51 points at its default level, as much as Google's top model Argon from September 30, which scored 50. But a task costs $1.87 instead of $4.97 and takes 7.9 instead of 30 minutes. What needed a top model six weeks ago now costs a third. Even so, Google is not the cheapest at this level: OpenAI's GPT-6.1 Sol gets there for $0.59.

What Google released

According to its model card, Hydrogen is adapted from Gemini 4 Carbon, Google's strongest model. Carbon and Argon remain the models for hard, long reasoning tasks. Hydrogen is built for fast execution: changing and testing code, searching documents, filling in spreadsheets, steps inside an agent. It reads text, images, audio, video and PDFs, has a context window of one million tokens and four thinking levels from "low" to "xhigh". Since November 12 it has been rolling out globally: in the developer API, the Gemini app, Google Search and Google Sheets.

Compared with the previous Gemini 3.8 Flash, Google reports big jumps: 2.5 times faster, 57 instead of 18.2 percent on the Terminal-Bench coding test, more than five times fewer fabricated claims. Also new is an Orchestrator in Antigravity, Google's coding environment. It switches between Carbon and Hydrogen within one session. According to Google it came close to the success rate of Carbon alone while using 24 percent less quota and finishing 19 percent faster.

Our analysis: what a task costs at this level

Artificial Analysis has every model solve the same set of tests and combines the results into one index. It also measures what a single task costs and how long it takes. From this data we picked all major models that land at around 50 points at one of their levels.

Cost per task at around 50 points in the index, in dollars

Average cost and duration per task in the Artificial Analysis Intelligence Index, retrieved on November 13, 2026, at list prices. For each model the level closest to 50 points. Bars on the same scale.

Per token, Hydrogen is only a quarter cheaper than Argon, $1.50 instead of $2.00 for input. Per task it is 62 percent cheaper. Most of the saving therefore does not come from the price tag but from the model getting to the result faster. From December 1 the gap grows further: Google then doubles the price of Argon and Carbon to $4 and $20.

The chart also shows how little the price per token tells you. GPT-6.1 Sol costs as much per token as Argon and is still the cheapest model per task in this comparison, more than three times cheaper than Hydrogen. If you choose models by the price list, you often choose the wrong one.

The most important dial: the thinking level

With Hydrogen you can set in four levels how much the model thinks. The measurement shows what each level delivers and costs.

Gemini 4 Hydrogen: points and cost per task by thinking level

Points in the indexCost per task

Artificial Analysis, Intelligence Index, retrieved on November 13, 2026. Points and cost each have their own scale, true to scale within each column.

The step from "low" to "medium" adds 14 points for 51 cents more. The last step from "high" to "xhigh" adds only 3 points but costs 69 percent more and takes 73 percent longer. Google itself warns in the model card that the model uses more tokens at higher levels. For most tasks the default level is the sensible choice, and "medium" is often enough for simple routine work.

The catch

  • Google's table is Google's table. In the model card Google compares Hydrogen with Claude Sonnet 5.5 and GPT-6.1 Terra, not with GPT-6.1 Sol, which is the cheapest at this level. Even there Sonnet 5.5 leads in three of ten tests, including operating a computer and coding in the terminal.
  • It is not the top. The strongest model in the index, Claude Fable 5.2, scores 64.6 points. For the hardest tasks Google itself still recommends Carbon and Argon.
  • The safety report is still missing. Google announces it for "the coming days". According to the model card, Hydrogen refuses harmless or borderline requests more often than 3.8 Flash, by 5.57 percentage points. Its knowledge ends in March 2026, and Google mentions occasional slowness and timeouts.
  • Free means Google may use your inputs. On the free tier Google uses the content to improve its products, on the paid tier it does not.

Google's own table: Hydrogen vs. Claude Sonnet 5.5, in percent

Gemini 4 HydrogenClaude Sonnet 5.5

Values from Google's model card, November 2026, measured by Google itself. Google measures the tenth test, Vending-Bench 2, in dollars instead of percent: $14,549 vs. $11,165 for Hydrogen. Bronze marks the three tests in which Sonnet 5.5 leads.

What this means for mid-sized companies

The level that needed Google's best model in September is mid-range in November. For many business tasks that is enough: reading offers and contracts, analyzing spreadsheets, summarizing research, single steps inside an agent. If you use Argon, Sonnet 5.5 or Opus 5.5 at a comparable level for this today, you pay roughly two to three times as much per task.

At the same time Google is clearing up. If an application still calls "gemini-2.5-flash" or "gemini-2.5-pro", it will automatically get a different model from December 1: 3.5 Flash-Lite instead of 2.5 Flash, 3.8 Flash instead of 2.5 Pro. Nothing changes in the code, but the answers may. Argon and Carbon cost twice as much from December 1, 3.8 Flash from January 1. For new projects Google explicitly recommends Hydrogen or 3.5 Flash-Lite.

What you should do now

  1. Count your Gemini calls. Where do your tools still use a model from the 2.5 series? Test the replacement with real cases before December 1, not after.
  2. Work out the December bill in advance. If you use Argon, Carbon or 3.8 Flash, you will soon pay double. Check where Hydrogen is enough.
  3. Compare by cost per task. Send 20 to 50 real tasks through two or three models and record result, cost and duration. The price per token is misleading.
  4. Set the thinking level on purpose. The default for most cases, one level lower for routine, the highest only where you can measure the gain.
  5. Let a big model plan and a fast one work. A strong model breaks the task down, a fast one does the parts, a person checks the result. That is how Google saves a quarter in its own Orchestrator.
  6. Company data only on a paid tier. The free tier is for trying things out, not for customer data.

Prices for a fixed level now fall within weeks, not years. If you recheck your choice of model once a quarter with your own tasks, you save more than any discount would. If you want to know which model handles your tasks at the lowest cost, we are happy to measure it with you on your own cases.

A conversation worth having, even without an engagement.

20 minutes, no obligation. Usually a personal reply the same working day.