#Gemini 4 Hydrogen#Google Gemini#xhigh#inteligencia artificial

    Gemini 4 Hydrogen thinking levels and the real cost

    Gemini 4 Hydrogen thinking levels: a task costs $0.29 to $3.16 depending on the level. How to set it and which one to start with.

    Gemini 4 Hydrogen thinking levels: Google made Gemini 4 Hydrogen generally available on November 12, 2026, at $1.50 per million input tokens and $7.50 per million output tokens, and its fourth reasoning level, xhigh, is what decides your bill. In one independent measurement, the same task costs $0.29 or over $3 depending on the level you pick. This guide explains what changes at each level, three uses with their monthly bill, and which level to start with.

    Analytics dashboard on a tablet: Gemini 4 Hydrogen's cost per task depends on the thinking level

    What changed with Gemini 4 Hydrogen

    Almost everyone covers it as another cheap model, but the news is not the price per token: it adds a fourth reasoning level, xhigh, alongside low, medium and high (the default). Only Hydrogen has xhigh; Gemini 4 Carbon and Argon have low, medium and high.

    Other official facts from Google:

    • It is on the free tier too.
    • Today it costs double per token what Gemini 3.8 Flash does, which stays at $0.75 input and $3.75 output through December 31, 2026. In January 2027 both cost the same.
    • It is the default model for the Antigravity agent in Gemini Managed Agents.

    How it works under the hood: you pay for all the thinking

    Hydrogen thinks before it answers, and that thinking has a cost. Reasoning tokens are billed as output, the expensive price, $7.50 per million. The API only gives you back a summary of that reasoning, but Google charges for all of it. On top of that, input, reasoning and answer share the same window of roughly one million tokens, and the output cap you set includes the reasoning.

    Step by step: how to set the thinking level

    With an example from Google's documentation, auditing a payment retry queue:

    1. Go to Google AI Studio, create an API key and pick the model gemini-4-hydrogen.
    2. Send the task and, inside generation_config, write thinking_level with low, medium, high or xhigh. If you write nothing, it uses high.
    3. Read the counter total_thought_tokens in the response: it tells you how much it thought and how much you are about to pay.
    4. Mind the output cap: if you set xhigh with a short cap, the reasoning eats the budget and the status comes back incomplete, with the answer cut off or empty. Raise the cap.
    5. Compare: run the same task on all four levels and check answer and cost.

    The bill by level: from $0.29 to $3.16 per task

    According to Artificial Analysis (cited by the-decoder.com; we did not read the original report), the cost per task and the Intelligence Index score are:

    | Level | Cost per task | Points | |---|---|---| | low | $0.29 | 27 | | medium | $0.80 | 41 | | high | $1.87 | 51 | | xhigh | $3.16 | 54 |

    Our analysis: from low to medium you pay $0.51 more and gain 14 points; from medium to high, $1.07 more and 10 points; from high to xhigh, $1.29 more (69% extra) and just 3 points. Each point costs about 1 cent at low and almost 6 at xhigh. The jump that pays off most is low to medium; the one that pays off least is the last. Also, xhigh needs almost double the output tokens of high.

    Code on screen: on Gemini 4 Hydrogen, reasoning is billed as output

    Three uses with their bill (our own analysis)

    These numbers are our own analysis with assumptions shown, not measured results.

    1. Sorting customer emails. An agency that sorts 500 emails a day (urgent, quote or complaint) handles 15,000 a month. With 1,000 input tokens and 500 output tokens with reasoning per email, at the low level: input $22.50 and output $56.25, total $78.75 in standard mode. In batch mode, which charges half, $39.38 if the report isn't urgent.

    2. Asking the same manual over and over. A 200,000-token manual queried 100 times a day adds up to 20 million input tokens: $30 without caching. With caching, hits cost $0.075 per million and the bill drops to about $1.79. On Hydrogen caching is implicit only: you don't turn it on and you don't control it; Google applies it when the beginning of your prompt repeats. Put the manual first and your question last.

    3. A hard-to-find bug. Two attempts on high cost $3.74, more than a single attempt on xhigh, $3.16. If the problem is hard, xhigh is cheaper than retrying. But that requires knowing up front when high won't be enough: the model no longer makes that call, you do.

    Should you migrate from Gemini 3.8 Flash?

    With the same tokens as the email case, Flash costs $11.25 for input and $28.12 for output through December 31: about $39.38, the same as Hydrogen in batch mode. Google says Hydrogen finishes tasks 2.5 times faster and replaces Flash, but that figure is Google's. If Flash already works for you, stay until December; in January they cost the same, and that is when to switch.

    Antigravity and agents

    Hydrogen is now the default model for Antigravity, Google's agent in Gemini Managed Agents. Every step that agent thinks is billed as output, and the documentation we checked does not say which level it uses by default. Before leaving it working alone all night, test it with a short task and look at how many reasoning tokens it used.

    Who it is not for

    • Anyone handling customer data on the free tier: Google may use that traffic to improve its products, so no contracts, histories or personal data go there.
    • Anyone who needs web search with the model: Search and Maps are not on the free tier; on paid they come with 5,000 free queries a month, then $14 per thousand.

    The objection: "just set xhigh and forget it"

    They are right about one part: per token it is cheap. But xhigh uses almost double the output tokens of high, and Google warns it can overthink simple tasks and exhaust the cap. You pay 69% more for three points. It wins when a mistake costs more than the bill: a payment retried wrong, a contract misread. For everything else, start on medium and only go up if it fails.

    What nobody is telling you

    The claim going around: Hydrogen beats Argon and nearly matches the big ones. The speed figures, the hallucination figures and the 57% Terminal-Bench number are published by Google about itself, and that last one is flagged as cheat-adjusted. The independent measurement from Artificial Analysis gives 51 points on high against 50 for Argon: that is even, not better, and one point below Claude Opus 5.5 and GPT-6.1 Sol. On terminal tasks it scores 54% against 64% for Claude Sonnet 5.5. What is true: from December, it costs less than half what Argon does. To believe all of it, run your own twenty tasks on the four levels.

    What really breaks

    1. Overthinking: on a simple question, xhigh can burn the output cap thinking and hand you back an empty answer, and you already paid for the reasoning.
    2. Implicit caching: change one comma at the start of the prompt and you lose the discount.
    3. Free-tier limits: Google publishes no numbers for this model, so don't build a business on it.

    How to apply it in your business?

    1. Start on medium, not the default level: high costs $1.87 against $0.80.
    2. Go up to high only if medium fails your test.
    3. Save xhigh for hard problems where a mistake costs more than $3.
    4. Set a high output cap and check that the status doesn't say incomplete.
    5. If the work isn't urgent, use batch and pay half.

    Track total_thought_tokens from the first week, because the thinking is your bill. For the short version of the story, read the full breakdown of Gemini 4 Hydrogen pricing.

    FAQ

    What are Gemini 4 Hydrogen's thinking levels?

    There are four: low, medium, high (the default) and xhigh. The higher the level, the more it reasons, and that reasoning is billed as output.

    How much does Gemini 4 Hydrogen cost?

    $1.50 per million input tokens and $7.50 per million output tokens. Batch and Flex cost half. According to Artificial Analysis, a task costs from $0.29 on low to $3.16 on xhigh.

    Which level should I use?

    Google recommends evaluating all levels and picking the lowest that works. Our rule: start on medium and only go up if it fails.

    Is there a free tier?

    Yes, but Google may use that traffic to improve its products and publishes no limits for this model.

    Conclusion

    What changed with Gemini 4 Hydrogen thinking levels is not the model: it is that now you decide how hard it thinks and how much you pay. Pick the task in your business worth paying to think harder, and measure before moving up a level.