Twice Flash's Price Per Token, a Quarter of Its Cost Per Task

Google announced Gemini 4 Hydrogen on 12 November as “our daily driver for agentic workflows.” It fills the slot the Flash models used to fill: the fast, cheaper model you run most of the time, with Gemini 4 Carbon and Gemini 4 Argon above it for hard problems. The post makes three kinds of claims. It’s priced flat at “$1.50 per million input tokens and $7.50 per million output tokens” across its 1 million-token window. Against Gemini 3.8 Flash, it finishes tasks “2.5 times faster at its high default setting.” And it’s for everyone, on every surface from AI Studio to the Gemini app to Search.

I checked each one against Google’s pricing page, Google’s help pages and Artificial Analysis’s model pages, all read on 12 November. The price looks like a step up until you count tokens per task. The speed claim checks out. The “everyone” claim is true in the place developers look and thinner in the place most people would meet the model.

The price is Flash’s list price

Google’s Gemini API pricing page lists Hydrogen at $1.50 input, $7.50 output and $0.075 for cached input, with implicit caching only. Gemini 3.8 Flash, the model Google says Hydrogen replaces for most teams, is listed at “$0.75 through December 31, 2026” for input and $3.75 for output, then “$1.50 starting January 1, 2027” and $7.50. So Hydrogen costs twice what Flash costs today, and exactly what Flash will cost seven weeks from now.

The models above it have their own dated step. Carbon and Argon are both listed at $2.00 input and $10.00 output “through November 30, 2026,” then $4.00 and $20.00 from 1 December. I wrote about these introductory prices when they first appeared in October. Put the three on one timeline and the comparison you get depends on the date you pick.

Output price per million tokens for Gemini 3.8 Flash, Gemini 4 Hydrogen and Gemini 4 Argon, 12 November 2026 to 31 January 2027 Paid-tier output prices from Google's Gemini API pricing page, read 12 November 2026. Gemini 4 Argon, and Carbon at the same price: 10 dollars per million output tokens through 30 November, 20 dollars from 1 December. Gemini 4 Hydrogen: 7.50 dollars throughout. Gemini 3.8 Flash: 3.75 dollars through 31 December, 7.50 dollars from 1 January 2027, the same as Hydrogen. Input prices follow the same pattern at one fifth of the output price. OUTPUT PRICE · USD PER 1M TOKENS · GEMINI API PRICING PAGE, READ 12 NOV 2026 $0 $5 $10 $15 $20 12 Nov 1 Dec 1 Jan 2027 31 Jan Argon and Carbon $10.00 through 30 Nov 2026 — Google pricing page, 12 Nov 2026 Argon and Carbon $20.00 from 1 Dec 2026 — Google pricing page, 12 Nov 2026 Hydrogen $7.50 — Google pricing page, 12 Nov 2026 3.8 Flash $3.75 through 31 Dec 2026 — Google pricing page, 12 Nov 2026 3.8 Flash $7.50 from 1 Jan 2027 — Google pricing page, 12 Nov 2026 Argon, Carbon $10 Argon, Carbon $20 Hydrogen $7.50 3.8 Flash $3.75 3.8 Flash joins Hydrogen at $7.50 Paid tier, standard (not batch). Dashed line: 3.8 Flash. Input prices are one fifth of output on all three.
Hydrogen's price doesn't move. Everything around it does: Argon and Carbon double on 1 December, and Flash doubles on 1 January to land on Hydrogen's line.

So the per-token comparison has three answers. Today Hydrogen is twice Flash and three quarters of Argon. From 1 December it’s 37.5% of Argon. From 1 January it’s the same as Flash.

Per task, it’s the cheaper one

A per-token price is half of a bill. The other half is how many tokens a job takes, and that can move the total by more than the rate does. A 2024 study of language models on hospital notes found up to a 17-fold cost reduction from packing 50 tasks into one prompt, at unchanged prices (Klang et al., 2024). I made a related point about cache reads in Opus and GPT-6 Sol bills.

Artificial Analysis publishes the per-task number for every model it runs through its Intelligence Index, a set of nine evaluations. It weights each evaluation’s cost by its weight in the index, so its totals can’t be rebuilt exactly from token counts, but the comparison across models uses the same method. On its pages for each model and thinking level:

At double the token price, Hydrogen at its default costs about a quarter of what Flash does per task: $1.87 against $7.77, a factor of 4.2. Most of that comes from reading about half as many input tokens. Google’s post says the same thing in its own words, that Hydrogen batches “file-read and search calls in parallel rather than inspecting files one by one,” and that partners saw token use fall by 12% to 23% against 3.8 Flash. Artificial Analysis’s run shows a larger drop than that on input.

Artificial Analysis Intelligence Index score against cost per index task, Gemini 4 Hydrogen at four thinking levels, Gemini 3.8 Flash and Gemini 4 Argon From Artificial Analysis model pages read 12 November 2026. Gemini 4 Hydrogen low: score 27, 0.29 dollars per task. Hydrogen medium: 41, 0.80 dollars. Hydrogen high, the default: 51, 1.87 dollars. Hydrogen xhigh: 54, 3.16 dollars. Gemini 3.8 Flash high: 31, 7.77 dollars, priced at its current introductory rate. Gemini 4 Argon high: 50, 9.95 dollars, priced at its 1 December list rate. Every Hydrogen setting costs less per task than Flash at high, and Hydrogen at medium already scores higher. INTELLIGENCE INDEX VS COST PER TASK · ARTIFICIAL ANALYSIS, READ 12 NOV 2026 20 30 40 50 60 $0 $2.50 $5 $7.50 $10 Hydrogen low: 27, $0.29 — Artificial Analysis, 12 Nov 2026 Hydrogen medium: 41, $0.80 — Artificial Analysis, 12 Nov 2026 Hydrogen high (default): 51, $1.87 — Artificial Analysis, 12 Nov 2026 Hydrogen xhigh: 54, $3.16 — Artificial Analysis, 12 Nov 2026 Gemini 3.8 Flash high: 31, $7.77 — Artificial Analysis, 12 Nov 2026 Gemini 4 Argon high: 50, $9.95 — Artificial Analysis, 12 Nov 2026 Hydrogen low · 27 medium · 41 high (default) · 51 xhigh · 54 3.8 Flash high · 31 Argon high · 50 Flash is priced at its intro rate ($0.75/$3.75) and Argon at its 1 Dec rate ($4/$20), as Artificial Analysis lists them.
Every Hydrogen setting sits to the left of Flash, and from medium up it also sits above it. Argon at high costs 5.3 times Hydrogen's default for one point less.

The comparison is uneven in one way. Artificial Analysis prices each model at the rate it lists today. Flash’s $7.77 uses the introductory $0.75 and $3.75. Every component of Flash’s price doubles on 1 January, so at list price the same run would cost $15.54 per task, a bit over eight times Hydrogen’s default. Argon’s $9.95 is already at the 1 December rate of $4 and $20. At the rate in force this month, half of that is $4.98, still more than two and a half times Hydrogen at high. Google’s headline that Hydrogen “surpasses Argon” on this index rests on one point, 51 to 50. The cost gap between them is 2.7x to 5.3x, depending on which Argon price you use.

The speed claim holds

Google’s “2.5 times faster” comes with its own conditions: high thinking against 3.8 Flash, on Google’s tasks. Artificial Analysis reports time per index task, which it computes as output tokens divided by measured end-to-end output speed. Flash at high takes 14.8 minutes per task. Hydrogen at high takes 6.0. That’s 2.47 times faster, on an outside run Google didn’t design.

Minutes per Artificial Analysis Intelligence Index task, by model and thinking level Time per Intelligence Index task from Artificial Analysis model pages, read 12 November 2026. Gemini 4 Hydrogen low: 1.3 minutes. Hydrogen medium: 3.7. Hydrogen high: 6.0. Hydrogen xhigh: 13.0. Gemini 3.8 Flash high: 14.8. Gemini 4 Argon high: 30.2. Flash at high divided by Hydrogen at high is 2.47. MINUTES PER INDEX TASK · LOWER IS FASTER · ARTIFICIAL ANALYSIS, READ 12 NOV 2026 Hydrogen low Hydrogen medium Hydrogen high (default) Hydrogen xhigh 3.8 Flash high Argon high 1.3 min — Artificial Analysis, 12 Nov 2026 3.7 min — Artificial Analysis, 12 Nov 2026 6.0 min — Artificial Analysis, 12 Nov 2026 13.0 min — Artificial Analysis, 12 Nov 2026 14.8 min — Artificial Analysis, 12 Nov 2026 30.2 min — Artificial Analysis, 12 Nov 2026 1.3 3.7 6.0 13.0 14.8 30.2 Time per task is output tokens divided by measured end-to-end output speed. 14.8 ÷ 6.0 = 2.47.
Google said 2.5x against Flash at the default setting. The outside number is 2.47x. At xhigh the advantage is mostly gone, because the model writes nearly three times as many tokens.

Two qualifiers. Time per task isn’t wall-clock time in your product, since it leaves out tool calls and your own latency, and the first answer token at high still takes about 15 seconds on Artificial Analysis’s measurement. And the speed lives at the default. Turn on xhigh for the extra three index points and Hydrogen takes 13.0 minutes per task, close to Flash.

Who gets it for free

The launch post says Hydrogen brings Gemini 4 “to everyone,” and lists the Gemini API, AI Studio, the Gemini app and Google Search. The pricing page backs up the developer part more than I expected. Hydrogen’s free-tier column reads “Free of charge” for input and output. Carbon and Argon are both marked “not available on the Free tier.” As before, free API use carries the condition “Used to improve our products: Yes.”

The app is different, and the post says so in a footnote: “free users will get access to Gemini 4 Hydrogen via auto model selection when it’s the appropriate model for the prompt.” Google’s help page on Gemini models spells out what that means for someone with no Google AI plan. The model column reads “Auto or Flash-Lite,” and the note under the table says “Most prompts will use Flash-Lite for quick answers. Prompts that require more reasoning may use Hydrogen. If auto model selection is turned off, all responses will use Flash-Lite.” Plus, Pro and Ultra subscribers can pick Hydrogen by name.

How a person without a paid plan reaches Gemini 4 Hydrogen, by surface From Google's Gemini API pricing page and Gemini Apps help page, read 12 November 2026. Gemini API free tier: Hydrogen is free of charge for input and output, and free-tier prompts are used to improve Google's products; Carbon and Argon are not on the free tier. Google Search: Hydrogen is rolling out for all users. Gemini app with no plan and auto model selection on: most prompts go to Flash-Lite, and prompts that need more reasoning may use Hydrogen; the user cannot pick it. Gemini app with no plan and auto selection off: all responses use Flash-Lite. Gemini app with Plus, Pro or Ultra: Hydrogen can be picked by name. WHERE A NON-PAYING USER MEETS HYDROGEN · GOOGLE PRICING AND HELP PAGES, 12 NOV 2026 Gemini API, free tier Google Search Gemini app, no plan, auto on Gemini app, no plan, auto off Developer surface Consumer surface Consumer surface Consumer surface Yes, by name Yes, rolling out Sometimes No Free of charge, input and output. Prompts used to improve products. Carbon and Argon: not on free tier. Launch post: upgrades Search for all users globally, with no model picker to choose it. "Most prompts will use Flash-Lite." Reasoning-heavy prompts "may use Hydrogen." The user can't choose. "All responses will use Flash-Lite." Plus, Pro and Ultra can pick Hydrogen by name. Quoted text is from Google's Gemini Apps help page and the launch post footnote.
"Everyone" is accurate for developers on the free API tier and for Search. In the free Gemini app it means Hydrogen when Google's router decides a prompt needs it.

That’s the same pattern I saw in October, when the free line in the consumer app moved down one model while the API kept a free Flash tier. This time the API free tier gets the newer model by name. A free app user gets it only when the router decides to send a prompt there.

What the model card defers

One more thing for anyone who reads model cards. The Hydrogen model card says “Gemini 4 Hydrogen is adapted from Gemini 4 Carbon,” and then repeats that sentence in place of an answer for architecture, training data, data processing, hardware and software, each pointing to Carbon’s card. That’s a reasonable disclosure for a derived model, and it confirms Hydrogen was built from Carbon rather than as a new Flash generation. It also leaves the card short on facts that belong to Hydrogen alone. The launch post’s safety numbers, such as the 96.7% honesty rate on “our internal agent honesty evaluation,” come from Google’s own tests and I couldn’t check them.

What I’d take from it

If you run Gemini 3.8 Flash today, Hydrogen at its default setting costs more per token and, on the one outside run published so far, less per task, while running about 2.5 times faster and scoring 20 points higher on the index. On 1 January the per-token difference goes away. Run your own traffic before you trust the factor of four, since how many tokens a model reads depends on your prompts and tools. The doubled rate card is the first number most teams will see, and on this evidence it overstates what they’ll pay.

If you were weighing Argon for everyday coding, the index has Hydrogen at high one point ahead for about a fifth of the per-task cost at Argon’s December price. And if you’re a free user in the Gemini app, you’ll get Hydrogen when Google’s router picks it, which it says won’t be most of the time.

References

  1. Kavukcuoglu, K. (2026). Introducing Gemini 4 Hydrogen. The Keyword, Google, 12 November 2026.
  2. Google (2026). Gemini Developer API pricing. Google AI for Developers, read 12 November 2026.
  3. Google (2026). Gemini 4 Hydrogen model card. Google DeepMind, November 2026.
  4. Google (2026). Gemini models in Gemini Apps. Gemini Apps Help, read 12 November 2026.
  5. Artificial Analysis (2026). Model pages for Gemini 4 Hydrogen (xhigh), Hydrogen (high), Hydrogen (medium), Hydrogen (low), Gemini 3.8 Flash (high) and Gemini 4 Argon (high). Intelligence Index v5.0.1, read 12 November 2026.
  6. Klang, E., Apakama, D., Abbott, E. E., Vaid, A., Lampert, J. et al. (2024). A strategy for cost-effective large language model use at health system-scale. npj Digital Medicine, 7, 320.