Skip to main content

Gemini 4 Hydrogen: flat pricing and cost per request

November 16, 20265 min. read
Gemini 4 Hydrogen: flat pricing and cost per request

On 12 November 2026 Google DeepMind presented Gemini 4 Hydrogen at US$ 1.50 per million input tokens and US$ 7.50 per million output tokens, with the same rate across the entire 1 million token window. With no tier by position, the cost of a request depends only on volume.

Key facts box Event: Gemini 4 Hydrogen, presented on 12 November 2026 by Google DeepMind. Rate: US$ 1.50 per million input and US$ 7.50 per million output, flat across the whole 1 million token window. Source: official announcement, consulted on 16 November 2026.

What Google fixed in Gemini 4 Hydrogen on 12 November

Google fixed the price of Gemini 4 Hydrogen at US$ 1.50 per million input tokens and US$ 7.50 per million output tokens, with no variation by token position inside the 1 million window. The announcement is dated 12 November 2026 and the rule holds for the full length of the window.

The official Hydrogen announcement presented the model as the daily-use one for agentic flow, with the page updated on 13 November 2026. The declared availability covers Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, Antigravity for paid subscribers and Google Sheets. For anyone arriving at the family now, the timeline includes Gemini 4 Carbon's restricted access for cyberdefense on 30 October 2026.

The window runs in both directions: native input of 1 million tokens and up to 1 million output tokens, with four reasoning levels, low, medium, high as the default and a new xhigh.

Price and invoice are different objects. A reader who stops halfway through the announcement concludes that the model got cheap. The announcement delivers a predictable pricing rule, and the distance between those two readings decides whether it is worth touching a route that already works.

What flat pricing across 1 million tokens stops penalizing

Under a flat price, the first input token and the millionth cost the same: US$ 1.50 per million. Cutting context to escape a more expensive tier stops having any effect, because there is no tier, and cost grows again only with the volume sent.

A 400,000 token prompt equals 0.4 million tokens at the input rate, or US$ 0.60, and the arithmetic does not change whether those 400,000 tokens sit at the start or at the end of the window. The cost curve of a request is a straight line, with no step anywhere.

The habit of truncating documents came from two reasons: fitting the window and spending fewer tokens. The first still holds. So does the second, because fewer tokens is less money at any rate. What falls away is the hope that cutting context changes where the token sits in the bill. In Hydrogen there is no position tier to escape.

The decision flat pricing returns to the team is which information has to be in the prompt, because every extra thousand tokens has a known price and is charged once, with no jump as context grows. The output window is the opposite axis of this piece, and Gemini 4 Argon and the 1M token output window already covered it. Here the question is what the position of the input token does to the bill.

inline-01.png

Figure: cost of one request with output fixed at 8,000 tokens, from 50,000 to 1 million input tokens, at both rates in the family. The lines are straight and the distance between them does not change. Source: Gemini 4 Hydrogen announcement, 12 November 2026.

The per-request bill under flat pricing

Three requests show how a flat price becomes dollars: US$ 0.66 for 400,000 input and 8,000 output, US$ 0.1125 for 60,000 and 3,000, which rounds to US$ 0.11, and US$ 0.00675 for 2,000 and 500. At the US$ 2 and US$ 10 rates of Carbon and Argon, Hydrogen costs 25% less.

RequestInputOutputInput costOutput costTotal on HydrogenTotal on Carbon or Argon
A, large prefix400,0008,000US$ 0.60US$ 0.06US$ 0.66US$ 0.88
B, multi-turn60,0003,000US$ 0.09US$ 0.0225US$ 0.1125US$ 0.15
C, high volume2,000500US$ 0.003US$ 0.00375US$ 0.00675US$ 0.009
  1. Input. Multiply input tokens by US$ 1.50 and divide by 1 million. In request A, 400,000 tokens give US$ 0.60.
  2. Output. Multiply output tokens by US$ 7.50 and divide by 1 million. In the same request, 8,000 tokens give US$ 0.06.
  3. Add the two lines. Request A closes at US$ 0.66.
  4. Repeat at the reference rate. At Gemini 4 Carbon at US$ 2 and US$ 10 per million, the same request gives US$ 0.88, and Hydrogen lands 25% below that number.
  5. Change the profile. Request C costs US$ 0.00675 and, at 10,000 requests a day, US$ 67.50.

The bill changes hands. In request A, input accounts for more than 90% of the cost; in C, output passes half of the total. The 25% distance holds while the reference rate is US$ 2 and US$ 10: the footnote in the Carbon announcement records the shift to US$ 4 and US$ 20 after 30 November 2026, declared valid only for public-access Carbon.

inline-02.png

Figure: the three requests in stacked bars, split into input and output, with the total at the tip. In request A input dominates; in C output passes half. Source: arithmetic on the announced rate, 12 November 2026.

Why token price stopped comparing models

Dollars per million compares rates, not requests. Two models with the same table deliver different costs when the output profile and the reasoning level change, and simulating one request carries the token mix the operation actually sends.

Google cites the Artificial Analysis Intelligence Index at 51 for Hydrogen at the default high level, against 50 for Argon, and 54 at xhigh. An index is not an invoice. The index is published by artificialanalysis.ai and appears here as Google cites it. The index page was not read for this check.

The same announcement states that Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at default high, more than triples the Terminal-Bench 4.1 result, at 57% against 18.2%, and cuts fabricated claims by more than 5 times. In early tests with partners, Google reports a gain of 2 to 2.8 times in speed and a drop of 12% to 23% in token use against the same 3.8 Flash. With Orchestrator in Antigravity, Google reports a success rate close to Carbon's on Terminal-Bench 4.0, with 24% less quota on average and 19% faster.

Quota is not dollars. None of those percentages becomes invoice savings without the load arithmetic, and this text does not make the conversion.

At the xhigh level the index rises from 51 to 54, and cost per request grows with it, because more reasoning comes out as more output tokens. The announcement does not state how reasoning tokens are billed, and the gap stays on the record.

Reusing context versus resending everything

Resending the same prefix on every turn now has a known ceiling: the prefix costs US$ 1.50 per million every time it enters the prompt. Keeping context in cache only beats that bill if the cached input price is lower, and the Hydrogen announcement does not state that price.

The resend path is the only one of the two with a closed price. Anyone running an agentic flow with a 400,000 token prefix over ten turns pays input ten times, which comes to 4 million input tokens and US$ 6.00 at the Hydrogen rate, and flat pricing allows calculating that cost before the first line of code is written. Doing the same with cache requires a number the announcement does not publish.

Without the cache price, the comparison between holding the prefix and resending the same 400,000 tokens has one measured side and one empty side, and deciding with half a number is choosing architecture by faith, not by arithmetic. Estimating that price by analogy with another version of the family would be inventing a number. The real cost of an AI application beyond the token price is where that bill goes when the data appears.

One number is missing.

The three-scenario test before switching models

The test that decides a switch measures three profiles: large prefix with medium output, multi-turn with a repeated prefix, and high volume with short output. In each one, record input tokens, output tokens and cost per completed task, then compare against the model in production before touching the route.

Scenario 1 reproduces request A and is the one flat pricing favors, because input dominates the cost and its price does not change with position. Scenario 2 measures how much the repeated prefix weighs, and it is where the missing cache price hurts most, because every turn resends the same block. Scenario 3 inverts the bill, with output passing half of the unit cost, and it is where flat pricing says least, since input price does not account per completed task.

Measure before switching.

For agentic flow with a large prefix, Hydrogen is the first option in the Gemini 4 family to enter a switching test, because flat pricing removes the most common reason to cut context. For a short request at very high volume, the decision stays open, and the criterion is cost per completed task, not price per million. Since the rate is the same anywhere in the window, the result of each scenario is comparable to the others, and the variable that decides the difference is the token mix of the load.

Frequently asked questions

The four questions below cover what flat pricing includes, the limit of the window, the difference between a predictable price and a lower price, and what is still missing for a switching decision. Each answer stands on its own, without depending on the sections above.

Does flat pricing apply to every Gemini 4 Hydrogen request?

Yes, inside the 1 million token window. The announcement states that the model is billed at US$ 1.50 per million input and US$ 7.50 per million output across the full length of the window, with no tier by token position, and the announcement is dated 12 November 2026.

What happens above 1 million tokens?

The announcement declares native input of 1 million tokens and up to 1 million output tokens, and does not state the behavior above that: an error, truncation or another rate is not written anywhere. Anyone who needs more context has to confirm the limit in the documentation before sizing the load.

Is flat pricing the same as a lower price?

No. Flat pricing means the rate does not change with token position, not that a request comes out cheaper than in another model. In the Gemini 4 family, Hydrogen costs 25% less than Carbon and Argon while the reference rate is US$ 2 and US$ 10, and the shift to US$ 4 and US$ 20 after 30 November 2026 belongs to Carbon, per the note in the announcement.

What is still missing to decide the switch?

Four things: the cached input price, the billing rule for reasoning tokens, rate limits per tier and an independent latency measurement. Without those numbers, the per-request bill closes, but cost per completed task stays open.

Sources and references

The two sources below are the only ones that supply a number in this piece, and no measurement was made independently for this text: indices, benchmarks and percentages appear as the vendor publishes them. Nothing was measured here. Both queries were made on 16 November 2026.

Next step: measure your own route

The per-request bill fits in a spreadsheet. Cost per completed task does not, and that is what decides a switch, because it depends on accepted quality and on a number of attempts that no announcement publishes. Hydrogen enters a route test as an executor with a known price.

A large prefix and a long output have a closed price. What stays open belongs to the load.

Nexforce Router enters this decision as a routing instrument, not as an offer: the catalog, price and availability of Hydrogen in the Router were not verified for this check, and nothing here authorizes treating the model as already available in Nexforce's mesh. To operate more than one model in the same session, the reference is context limits and capacity when operating several models.

Start with the token log. It already carries the mix that no benchmark delivers.

Nexforce

Accelerate your company'sbusiness and operational efficiency

We design the technology of tomorrow to boost your business operational scale

Talk to a Specialist

Related articles