Google / News

Gemini 4 Hydrogen tested: free in the Gemini API, and every check passed

Google's new Gemini 4 Hydrogen is free on the Gemini API's free tier. On our secret coding test it passed all 186 checks; through the paid API a job cost about 7.6 cents on average.

Table: Gemini 4 Hydrogen passed 186 of 186 checks for about 7.6 cents per job in 69 seconds; Gemini 3.8 Flash 8.5 cents and 147 s, Gemini 4 Argon 1.4 cents and 24 s, Gemini 4 Carbon 14.9 cents and 172 s; all passed every check
Our secret coding test: Hydrogen and Carbon on 12 November 2026, Argon and Flash in October. Chart by Not an AI App.

On 12 November 2026 Google released Gemini 4 Hydrogen. It is a fast model for everyday coding, and its price per token is lower than that of Google's big models, Gemini 4 Carbon and Argon. It is also on the free tier of the Gemini API, the connection programmers use to build with Gemini. We gave it our secret coding test the same day. Hydrogen passed all 186 of 186 checks. Through the paid API, a small job cost about 7.6 cents on average, and on the free tier it costs nothing per token.

What is Gemini 4 Hydrogen?

Screenshot of the top of Google's post 'Introducing Gemini 4 Hydrogen', dated 12 November 2026
Our own capture of Google's blog, 13 November 2026, cropped only.

Gemini 4 Hydrogen is Google's new everyday model. The announcement says it brings "cutting-edge intelligence to everyone". Google's bigger models, Carbon and Argon, are meant for deep thinking on hard problems. Hydrogen is meant for fast work: writing code, running tests and checking changes.

Google brings it to "all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search". People with a paid Google AI Pro or Ultra plan can also use it in Antigravity, Google's coding tool, the developer post says.

Hydrogen has four "thinking levels": low, medium, high and a new xhigh. High is the default. A higher level means the model thinks longer before it answers. That can help on hard jobs, but it costs more tokens. A token is a small piece of text; AI companies count and bill text in tokens.

Is Gemini 4 Hydrogen free?

Table: on the free tier of the Gemini API, Gemini 4 Hydrogen is free of charge and requests are used to improve Google's products; on the paid tier it costs $1.50 in and $7.50 out per million tokens and requests are not used for that; limits are shown in Google AI Studio
From Google's price list and rate-limit page, read 12 and 13 November 2026. Chart by Not an AI App.

In the Gemini API, partly. Google's price list shows two tiers:

  • Free tier: input and output are "free of charge".
  • Paid tier: $1.50 per million tokens in and $7.50 per million tokens out. That price is the same for any length of text, up to one million tokens, the most the model can handle at once.

There is a catch. On the free tier, the price list says your requests are "used to improve our products". On the paid tier, they are not. So do not send private code or customer data on the free tier.

The free tier also has limits on how much you can send. Google's rate-limit page says your limits depend on factors such as your usage tier, and that you can view them in Google AI Studio. Google's bigger models, Gemini 4 Carbon and Argon, are not on the free tier at all.

Our test

Diagram of our test: three secret coding jobs, three tries each, sent to Gemini 4 Hydrogen through OpenRouter with no extra settings on 12 November 2026, graded with 186 hidden checks
Diagram by Not an AI App.

We gave Hydrogen the same secret coding test as other models. It has three jobs: fix the money maths in a small program, write a bill splitter, and write a CSV reader. A CSV reader turns a text file with commas into a table. Each job ran three times. One round of the three jobs has 62 hidden checks, so three rounds give 186 check results. The AI never sees those checks.

We ran Hydrogen through OpenRouter, a paid service that gives access to many models, on 12 November 2026. Each request sent only our task text, without extra settings. We did not test the free tier, because that needs a separate Google account and key.

What does one coding job cost?

Bar chart of the average cost per coding job: GPT-6 Luna 0.12 cents, Claude Haiku 5.5 0.19, GPT-6.1 Sol 1.3, Gemini 4 Argon 1.4, Claude Sonnet 5.5 2.6, Gemini 4 Hydrogen 7.6, Gemini 3.8 Flash 8.5 and Gemini 4 Carbon 14.9 cents
Our secret coding test through the paid API (Sol through OpenAI). Hydrogen and Carbon 12 November, the others 6 to 17 October 2026. Chart by Not an AI App.

Hydrogen passed all 186 checks. Through the paid API, a job cost about 7.6 cents on average. We compare it with earlier runs of the same test:

ModelChecks passedCost per jobTypical time
GPT-6 Luna184/1860.12 cents25 s
Claude Haiku 5.5186/1860.19 cents15 s
GPT-6.1 Sol186/1861.3 cents22 s
Gemini 4 Argon186/1861.4 cents24 s
Claude Sonnet 5.5186/1862.6 cents17 s
Gemini 4 Hydrogen186/1867.6 cents69 s
Gemini 3.8 Flash186/1868.5 cents147 s
Gemini 4 Carbon186/18614.9 cents172 s

Cost per job is the average of nine runs; typical time is the middle value. So Hydrogen is not cheap per job on the paid tier. Its price per token is low, but it thinks a lot. About 90% of its output tokens were reported as thinking tokens. You pay for them, even though they are not part of the answer you see. The middle value for its output tokens was about 10,300 per job, thinking tokens included. Claude Haiku 5.5 passed every check for a few tenths of a cent; GPT-6 Luna was even cheaper but missed two checks. Even Google's own Argon was cheaper per job, although its price per token is higher ($2 in and $10 out until 30 November). See our price comparison for more.

Hydrogen or Gemini 3.8 Flash?

Bar chart of typical time per job: Gemini 4 Hydrogen 69 seconds and Gemini 3.8 Flash 147 seconds; about 2.1 times as fast on our jobs; Google reports 2.5 times on its own tests
Our secret coding test. Chart by Not an AI App.

Google calls Hydrogen "a major standalone upgrade for teams migrating from Gemini 3.8 Flash", its older fast model. Google says Hydrogen finishes tasks "2.5 times faster" at its default setting.

On our jobs, Hydrogen took about 69 seconds as the middle value, against 147 seconds for Gemini 3.8 Flash. So on our jobs, Hydrogen was about 2.1 times as fast. Google reports 2.5 times on its own tests; those results are not directly comparable with ours. Both passed every check. Hydrogen wrote fewer output tokens: the middle value was about 10,300 per job, against 27,100 for Flash, thinking tokens included. So it was also a little cheaper: 7.6 against 8.5 cents.

Our Flash runs were on 10 October and our Hydrogen runs on 12 November, both through OpenRouter. Speed can change from day to day, so take the exact numbers lightly.

What Google says it can do

Screenshot of Google's post: Hydrogen is for all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search; it scores 51 on the Artificial Analysis Intelligence Index at high and 54 at xhigh, and costs $1.50 in and $7.50 out
Our own capture of Google's blog, 13 November 2026, cropped only.

Google shares many test scores. These are Google's numbers, on tests Google picked:

  • Artificial Analysis Intelligence Index, a mix of tests run by an outside company: 51 at the default high level, just above Argon's 50, and 54 at xhigh. The Artificial Analysis page agreed on 13 November: Hydrogen (Xhigh) scores 54.
  • Terminal-Bench 4.1, a test of coding jobs in a terminal, the window where you type commands: 57%, against 18.2% for Gemini 3.8 Flash, Google says.
  • Fewer made-up claims: Google says Hydrogen cuts "fabricated claims by more than 5x" compared with Gemini 3.8 Flash.

We could only check Hydrogen's xhigh score of 54 ourselves. The score of 51 at high, Argon's 50 and the other results are Google's claims. Our own small test fits the picture of a model that gets code right, but it says nothing about long projects.

Should you use it?

Overview: Hydrogen on the free tier for a free Gemini model if Google may use your data; Claude Haiku 5.5 for the cheapest code that passed every check; Gemini 4 Argon as the cheapest Google model per job we tested; Hydrogen instead of Gemini 3.8 Flash
Based on our test and Google's price list. Chart by Not an AI App.
  • You want a free Gemini model in the API? Hydrogen on the free tier, as long as Google may use what you send to improve its products.
  • You want the cheapest code that passed every check, and pay per token? On our test, that was Claude Haiku 5.5, at about 0.19 cents per job. GPT-6 Luna was cheaper still, but missed two checks.
  • You want the cheapest Google model per job among those we tested? Gemini 4 Argon, at about 1.4 cents per job. Google's price list says its price goes up from $2 and $10 to $4 and $20 on 1 December.
  • You still use Gemini 3.8 Flash? Hydrogen was about twice as fast and a bit cheaper on our jobs.
  • You want Google's model for long, hard jobs? Google points to Gemini 4 Carbon and Argon. On our small jobs, Carbon cost about 14.9 cents per job.

Questions people ask

Ten selected questions from Google's "People also ask" about the Gemini API and free AI APIs, with the 4 answered in this article marked
"People also ask" questions from Google (US) for "gemini api free tier", "gemini flash" and "free ai api", 13 November 2026, via DataForSEO. Chart by Not an AI App.

What does Gemini Free Tier offer?

On the free tier of the Gemini API, some models cost nothing per token, such as the new Gemini 4 Hydrogen. Gemini 4 Carbon and Argon are not on the free tier. Google's price list says free-tier requests are "used to improve our products"; paid ones are not. Your limits are shown in Google AI Studio.

Is Gemini free to use?

Partly. Google says Gemini 4 Hydrogen is available to everyone in the Gemini app and Google Search, and on the free tier of the Gemini API. Gemini 4 Carbon is not on the Gemini API's free tier.

Can Gemini do programming?

Yes. On our secret coding test of 12 November 2026, Gemini 4 Hydrogen passed all 186 of 186 checks, for about 7.6 cents per job on average through the paid API. Our jobs were small, so this does not show how it does on big projects.

Can I get a free AI API?

Yes, with limits. The Gemini API has a free tier, where Gemini 4 Hydrogen costs nothing per token. Google may use what you send there to improve its products. Check what happens with your data first.

How we checked

Diagram of how we checked: read Google's announcement, price list and rate-limit page, run nine coding jobs through OpenRouter, grade them with 186 hidden checks and take the cost from OpenRouter
Diagram by Not an AI App.

On 12 and 13 November 2026 we read Google's announcement and developer post, and the Gemini API update notes, price list and rate-limit page. On 13 November we checked the index score on Artificial Analysis.

We ran our secret coding test through OpenRouter on 12 November: three jobs, three tries each, 186 checks, one request per job. The costs are what OpenRouter charged; they match Google's list prices. Gemini 4 Carbon ran on the same jobs on 12 November; the other models between 6 and 17 October 2026. GPT-6.1 Sol ran through OpenAI itself. Its cost is worked out from OpenAI's token counts at list price. Our methods page explains the test.

The screenshots are our own captures of Google's pages, only cropped. The charts are our own, and none of the images was made by AI. Our jobs are small and clear; your results on big projects may differ.