# Gemini 4 Hydrogen vs Carbon and Argon: which Gemini model to choose

> Gemini 4 Hydrogen vs Carbon and Argon: roles, prices, reasoning, 3.8 Flash and Flash-Lite compared. Which Gemini model to use, with two cost examples.

Source: Giallo Studio (giallostudio.it) · Page: https://giallostudio.it/en/blog/gemini-4-hydrogen-vs-carbon-argon · Published: 2026-11-14 · Updated: 2026-11-14

**In the Gemini 4 Hydrogen vs Carbon comparison, the model Google presented on 12 November 2026 is for getting work done fast (code, tools, documents), while Carbon and Argon, the “frontier” models, stay for hard problems and long tasks.** The price follows this split: Hydrogen costs $1.50 per million input tokens, Carbon and Argon $4.00 from 1 December 2026. The data comes from the [official announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/), the guide [What’s new in Gemini 4 Hydrogen](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen), the [Gemini API price list](https://aistudio.google.com/docs/pricing) and the [Gemini app models page](https://support.google.com/gemini/answer/18648722), all accessed on 14 November 2026.

For an SME the practical question is “which model do I put in my workflow?” This guide compares the five reference text models in the price list: Hydrogen, Argon, Carbon, Gemini 3.8 Flash and Gemini 3.5 Flash-Lite. It follows only what Google recommends and reports; Giallo has not rerun the benchmarks. For the definition of the model, see the guide [Gemini 4 Hydrogen: what it is](https://giallostudio.it/en/blog/gemini-4-hydrogen-what-it-is).

<div class="key-points">
<p class="kp-title">In summary</p>
<ul>
<li>Google describes Hydrogen as the model for “rapid execution” and Carbon and Argon as “frontier” models for deep reasoning; Hydrogen is “adapted from Gemini 4 Carbon”, that is, derived from Carbon.</li>
<li>Hydrogen costs $1.50 for input and $7.50 for output per million tokens (<a href="https://aistudio.google.com/docs/pricing">Google price list</a>); Argon and Carbon go from $2.00 and $10.00 to $4.00 and $20.00 on 1 December 2026.</li>
<li>On the Artificial Analysis Intelligence Index, Google reports 51 for Hydrogen against 50 for Argon, and 54 at the highest reasoning level.</li>
<li>Compared with Gemini 3.8 Flash, Google reports tasks finished 2.5 times faster and 57% against 18.2% on Terminal-Bench 4.1.</li>
<li>Carbon and Hydrogen can work as a pair: in Antigravity's Orchestrator, on Terminal-Bench 4.0, Google reports 24% less quota used, finishing the work 19% faster.</li>
<li>Gemini 3.5 Flash-Lite remains, according to Google, the fastest and cheapest model for classification, routing and sub-second latency.</li>
</ul>
</div>

## How do Gemini 4 Hydrogen, Argon, Carbon, 3.8 Flash and Flash-Lite compare?

The table puts the five models on the same columns. Prices are in dollars per million tokens, from the [Gemini API price list](https://aistudio.google.com/docs/pricing); roles, reasoning levels and output limits come from the [Google guide to Hydrogen](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen) and the [model list](https://aistudio.google.com/docs/models); app plans come from the [Gemini app guide](https://support.google.com/gemini/answer/18648722), valid for personal Google accounts (with a work or school account they depend on the Workspace license).

| Model (API ID) | Role according to Google | Price today, input / output | Price after the change | Reasoning (default) | Output per step | API free tier | App plans |
|---|---|---|---|---|---|---|---|
| **Hydrogen** `gemini-4-hydrogen` | Fast execution: everyday code, agents with tools, multimodal documents, parallel subagents | [1,50 $ / 7,50 $](https://aistudio.google.com/docs/pricing) | No change indicated; flat price across 1M tokens | `low`, `medium`, `high` (default), `xhigh` | Up to 65,536 tokens per synchronous request | Yes | Plus, Pro, Ultra; automatic selection even without a plan |
| **Argon** `gemini-4-argon` | Advanced, long tasks with fewer tokens than Carbon | [$2.00 / $10.00 until 30 Nov 2026](https://aistudio.google.com/docs/pricing) | $4.00 / $20.00 from 1 Dec 2026 | `medium` (default), maximum `high` | 32,768 tokens per non-streaming step via REST | No | Pro, Ultra |
| **Carbon** `gemini-4-carbon` | The strongest model: architecture, debugging, science, security audits | [$2.00 / $10.00 until 30 Nov 2026](https://aistudio.google.com/docs/pricing) | $4.00 / $20.00 from 1 Dec 2026 | `medium` (default), maximum `high` | 32,768 tokens per non-streaming step via REST | No | Ultra |
| **3.8 Flash** `gemini-3.8-flash` | Previous Flash generation; Hydrogen is the recommended upgrade | [$0.75 / $3.75 until 31 Dec 2026](https://aistudio.google.com/docs/pricing) | $1.50 / $7.50 from 1 Jan 2027 | `medium` (default, per the migration guide); no `xhigh` | 64K output budget, separate from the 1M input | Yes | Not listed in the guide |
| **3.5 Flash-Lite** `gemini-3.5-flash-lite` | The fastest and cheapest: classification, routing, sub-second latency | [0,30 $ / 2,50 $](https://aistudio.google.com/docs/pricing) | No change indicated | Not specified on the pages consulted | Not specified on the pages consulted | Yes | “Flash-Lite” on all plans, version not specified |

Three notes on the table. Carbon, Argon and Hydrogen share the same single 1,048,576-token window and output of up to 1 million tokens with “long decode”, long generation in steps (Google [describes the differences between steps](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen)). Hydrogen's price is flat across the whole window: no pricier tier beyond a certain length. Carbon has been in the paid API since 10 November 2026, according to [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/).

The full price list, with Batch, Flex, Priority and cache, is in the sibling article [Gemini 4 Hydrogen: API pricing and costs](https://giallostudio.it/en/blog/gemini-4-hydrogen-api-pricing).

## What is the difference between Hydrogen and Argon?

According to Google, the main difference is one of role and price: Argon and Hydrogen share context, built-in tools and long decode. Google recommends Argon for “the majority of advanced long-horizon tasks” with fewer tokens than Carbon; it recommends Hydrogen for everyday programming, interactive web apps, agents with tools, multimodal documents and parallel subagents ([Google guide](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen)).

From 1 December 2026 the price gap becomes sharp: Argon costs $4.00 for input and $20.00 for output, Hydrogen $1.50 and $7.50, so Argon costs 2.67 times as much as Hydrogen on both lines ([price list](https://aistudio.google.com/docs/pricing)). Until 30 November 2026 the ratio is 1.33.

On the Artificial Analysis Intelligence Index, the announcement reports 51 for Hydrogen at the default level `high` against 50 for Argon, and 54 at level `xhigh` ([announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)). That is Google's own figure, on a synthetic index. On that index Hydrogen does not sit below Argon; the figure does not say it replaces Argon on every long task. For individual tests against Claude and GPT, see the article [Gemini 4 Hydrogen benchmarks explained](https://giallostudio.it/en/blog/gemini-4-hydrogen-benchmark); for the Carbon and Argon pair, see the comparison [Carbon vs Argon](https://giallostudio.it/en/blog/gemini-4-carbon-vs-argon).

One operational detail: Hydrogen uses `high` as its default reasoning level, while Carbon and Argon use `medium` with a maximum of `high`. For the same request, then, token consumption is not comparable until you set the same levels. How the levels work is explained in [reasoning effort: what changes](https://giallostudio.it/en/blog/reasoning-effort-explained).

## Why does Google pair Carbon and Hydrogen?

Because planning and building are two different jobs. Google suggests having Carbon plan a system's interfaces or solve a hard blocker, and delegating code changes, parallel subagents and tests to Hydrogen ([Google guide](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen)). In a multi-agent flow, Carbon can act as coordinator and Hydrogen as executor and verifier.

The feature is called Orchestrator and lives in Google Antigravity: it switches between models, for example Carbon and Hydrogen, within the same session. On Terminal-Bench 4.0 Google reports “near-Carbon” success rates, with 24% less quota used on average, finishing the work 19% faster ([announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)). Antigravity is available for Hydrogen to developers with a paid Google AI Pro or Ultra subscription; for businesses, the announcement lists it among the “For enterprises” channels, together with Gemini for business.

Anyone working through the API can replicate the idea by hand: the guide says to route planning turns to `gemini-4-carbon` and execution turns to `gemini-4-hydrogen` within the same session, using `previous_interaction_id`. For a software house, it pays to decide up front which steps need the most expensive model. Integration details are in the [migration checklist for Hydrogen](https://giallostudio.it/en/blog/gemini-4-hydrogen-migration-checklist).

## Hydrogen or Gemini 3.8 Flash: should you switch?

Google recommends it: “Recommended upgrade from Gemini 3.8 Flash and earlier Gemini 3 and 2.5 models” ([Google guide](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen)). The figures reported in the announcement, which Giallo has not verified, are these ([Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)):

- tasks finished 2.5 times faster at the default level `high`;
- Terminal-Bench 4.1 at 57% against 18.2%: a terminal coding test, different from the Terminal-Bench 4.0 used for the Orchestrator;
- fabricated claims reduced by more than 5 times;
- in early partner tests, 2-2.8 times faster and 12-23% fewer tokens on graphs, SQL and business data ([Google announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)).

On price the picture is less obvious. Until 31 December 2026, Gemini 3.8 Flash costs $0.75 for input and $3.75 for output, half of Hydrogen; from 1 January 2027 it costs $1.50 and $7.50, the same as Hydrogen ([price list](https://aistudio.google.com/docs/pricing)). From 2027, then, what remains is the difference in quality and features, not in price. One limit to know about: in the [model card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/) Google reports, compared with 3.8 Flash, unjustified refusals up by 5.57 points, mostly with multimodal prompts. If your integration handles images and documents, measure it.

The switch is not a rename: the context window, reasoning levels, caching and sampling parameters all change. The path is in the [migration checklist](https://giallostudio.it/en/blog/gemini-4-hydrogen-migration-checklist). If you are coming from Gemini 3.6 and 3.5, the previous comparison is [Gemini 3.6 Flash vs 3.5 Flash](https://giallostudio.it/en/blog/gemini-3-6-flash-vs-3-5-flash).

## Which Gemini model should you use for each task?

The table follows only Google's recommendations in the [Hydrogen guide](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen), in the [announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) and in the [price list](https://aistudio.google.com/docs/pricing). It contains no numeric thresholds, because Google does not publish any on the pages consulted on 14 November 2026.

| Task | Model recommended by Google | Note |
|---|---|---|
| Everyday programming, bug fixing, interactive web apps | Hydrogen | Level `high` is the default; `medium` for interactive code and responsive chat |
| Agent that uses tools, parallel subagents | Hydrogen | As the executor, even under a Carbon coordinator |
| Dense PDFs, tables, charts, audio, video to turn into reports | Hydrogen | Google cites it for multimodal document workflows |
| Hard architecture, tricky debugging, science, security audits | Carbon | Spends more reasoning tokens to check edge cases |
| Advanced, long tasks, with fewer tokens than Carbon | Argon | Same price as Carbon |
| Plan a complex job, then carry it out | Carbon and Hydrogen together | Carbon plans, Hydrogen executes |
| Hard problem where `high` is not enough | `xhigh` on Hydrogen | More tokens and latency; on simple prompts it can lead to “overthinking” |
| High-volume classification, routing, sub-second responses | 3.5 Flash-Lite | Google recommends it for these uses |
| Integration still on 3.8 Flash, Gemini 3 or 2.5 | Hydrogen | Recommended upgrade; test first |

The line between cases where Hydrogen at `high` is enough and cases where you need Carbon is not written down anywhere: you find it by testing your own cases, as in the last section.

## Which Gemini model should a marketing agency choose for client reports?

A worked example, not a real case. An agency produces 40 monthly reports for its clients. Assumption: each report processes 60,000 input tokens (exported data, extracts, briefs) and produces 12,000 output tokens, including reasoning tokens, which are billed as output ([price list](https://aistudio.google.com/docs/pricing)). Total: 2,400,000 input tokens and 480,000 output tokens per month, with no cache discount.

| Model (price list) | Monthly cost of the example |
|---|---|
| Hydrogen, [1,50 $ / 7,50 $](https://aistudio.google.com/docs/pricing) | 3,60 $ + 3,60 $ = **7,20 $** |
| Argon or Carbon until 30 November 2026, [2,00 $ / 10,00 $](https://aistudio.google.com/docs/pricing) | 4,80 $ + 4,80 $ = 9,60 $ |
| Argon or Carbon from 1 December 2026, [4,00 $ / 20,00 $](https://aistudio.google.com/docs/pricing) | 9,60 $ + 9,60 $ = 19,20 $ |
| 3.8 Flash until 31 December 2026, [0,75 $ / 3,75 $](https://aistudio.google.com/docs/pricing) | 1,80 $ + 1,80 $ = 3,60 $ |
| 3.5 Flash-Lite, [0,30 $ / 2,50 $](https://aistudio.google.com/docs/pricing) | 0,72 $ + 1,20 $ = 1,92 $ |

The figures are small: at these volumes the model matters for quality, not for cost. A report is synthesis work on documents, tables and charts, the ground Google assigns to Hydrogen. Flash-Lite costs less, but Google recommends it for classifying and routing: in the agency's flow it makes sense for triaging requests, not for writing the report. Argon or Carbon come in only if testing shows errors that Hydrogen at `high` or `xhigh` does not fix. For work with SMEs, see the guide [Gemini 4 Hydrogen for business](https://giallostudio.it/en/blog/gemini-4-hydrogen-for-business). The calculation method is covered in [AI API costs: how to calculate them](https://giallostudio.it/en/blog/how-to-calculate-ai-api-costs).

## Which Gemini model should a development studio choose for a coding agent?

Another worked example, with stated assumptions. A studio runs an agent on 200 tasks a month. Each task uses 300,000 input tokens accumulated across steps and 30,000 output tokens, reasoning included: 60 million input and 6 million output. Hydrogen's implicit caching, which discounts cached input tokens by 95%, is ignored, so actual spend may be lower ([Google guide](https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen)).

| Scenario | Calculation with the prices in the [price list](https://aistudio.google.com/docs/pricing) | Monthly cost of the example |
|---|---|---|
| [Hydrogen only](https://aistudio.google.com/docs/pricing) | 60 × 1,50 $ + 6 × 7,50 $ | **135 $** |
| [Argon or Carbon only](https://aistudio.google.com/docs/pricing) from 1 December 2026 | 60 × 4,00 $ + 6 × 20,00 $ | 360 $ |
| [Pair](https://aistudio.google.com/docs/pricing): 20% of tokens on Carbon, 80% on Hydrogen (assumption) | 12 × 4,00 $ + 1,2 × 20,00 $ + 48 × 1,50 $ + 4,8 × 7,50 $ | 180 $ |

The pair costs $45 more than Hydrogen alone, in this assumption with the prices in the [price list](https://aistudio.google.com/docs/pricing). It pays for itself only if it reduces failed attempts or human review: the 20% is a choice made in the example, not a Google figure. The Orchestrator reports quota savings compared with Carbon, but on Terminal-Bench 4.0 and inside Antigravity, not in the calculation above. To connect the agent to an existing system, the [software house](https://giallostudio.it/en/software-house) page at Giallo explains how we work; for agents with tools, see [AI agents](https://giallostudio.it/en/ai-agents).

<div class="compare">
<div class="compare-yes"><h4>Hydrogen at the high level is enough if</h4><ul><li>the work is execution: code, tools, documents, reports;</li><li>volumes are high and the cost per million tokens matters;</li><li>you need fast responses;</li><li>you are starting from 3.8 Flash and want the recommended upgrade.</li></ul></div>
<div><h4>Try Argon or Carbon if</h4><ul><li>the task is architecture, difficult debugging or scientific analysis;</li><li>Hydrogen at <code>xhigh</code> still gets edge cases wrong;</li><li>an error costs far more than the tokens spent;</li><li>a coordinator has to plan the work of several agents.</li></ul></div>
</div>

## How do you choose a Gemini model in one week, without trusting the headlines?

Google's benchmarks tell you where a model is strong on average, not how it performs on your documents. A five-step test is a good starting point.

<ol class="steps">
<li><strong>Collect 15-20 real cases that are already closed:</strong> with the correct result known, including two or three incomplete or ambiguous cases.</li>
<li><strong>Test Hydrogen at all four levels:</strong> Google recommends running the test at <code>low</code>, <code>medium</code>, <code>high</code> and <code>xhigh</code> with a <code>max_output_tokens</code> set high, so the limit does not cut off responses, and choosing the lowest level that reaches the required quality.</li>
<li><strong>Measure three things:</strong> errors that make the result unusable, minutes of human review, tokens consumed per case.</li>
<li><strong>Move up to Argon or Carbon only on the cases that fail:</strong> if the error disappears and the cost stays acceptable, you have a routing rule.</li>
<li><strong>Lock in the choice and retest at every price list change:</strong> 1 December 2026 for Carbon and Argon, 1 January 2027 for 3.8 Flash.</li>
</ol>

## What mistakes should you avoid when choosing between Gemini models?

- **Comparing costs at different reasoning levels.** Hydrogen starts at `high`, while Argon and Carbon start at `medium`: set them the same way before reading the bill.
- **Choosing the strongest model “just to be safe.”** At $4.00 and $20.00 from 1 December 2026, Carbon and Argon cost 2.67 times as much as Hydrogen: it makes sense where the task requires it.
- **Forgetting the dates.** The introductory price for Carbon and Argon ends on 30 November 2026; comparing a December quote with the November price list is a mistake.
- **Using the app to test the API.** The Gemini app has its own limits, plans and reasoning levels (Low, Medium, High, no xhigh); results do not translate one to one into API costs and behavior.
- **Treating a Google figure as universal truth.** Benchmarks are reported by the vendor, often with its own configurations ([methodology](https://deepmind.google/models/evals-methodology/gemini-4-hydrogen)).

## Where should an SME start when choosing the right model?

With testing, not with the model name. Hydrogen is the natural candidate for daily work; Carbon and Argon are the exception, to be justified by a case that fails; Flash-Lite is the cheap filter in front of everything. The choice stays valid only as long as the price list and availability do not change: always check the official pages cited.

If you want a test bench on your real cases, with Hydrogen's four reasoning levels and a look at costs from 1 December 2026, the [AI consulting from Giallo](https://giallostudio.it/en/ai-consulting) can start from one process, 15-20 examples and a week of measurements, and deliver a routing rule across the models.

## Frequently asked questions

### Is Gemini 4 Hydrogen better than Carbon?

No, not in absolute terms: they are built for different jobs. Google presents Carbon as its strongest model for deep reasoning and Hydrogen as the model for fast execution. The pages consulted on 14 November 2026 contain no benchmark table putting Hydrogen and Carbon side by side, so the choice comes down to a test on your own cases.

### Is Gemini 4 Hydrogen cheaper than Argon?

Yes. Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Argon costs $2.00 and $10.00 until 30 November 2026 and $4.00 and $20.00 from 1 December 2026, about 2.7 times Hydrogen. According to Google, Hydrogen scores 51 on the Artificial Analysis Intelligence Index against 50 for Argon.

### Is it worth switching from Gemini 3.8 Flash to Gemini 4 Hydrogen?

Google recommends it as the upgrade, with tasks finished 2.5 times faster and 57% against 18.2% on Terminal-Bench 4.1: these are figures reported by Google. Until 31 December 2026, 3.8 Flash costs half as much; from 1 January 2027 it costs the same as Hydrogen. Switching requires changes to the integration, so it needs testing.

### Which Gemini model should you use for a coding agent?

It depends on the task. According to Google, Carbon suits hard architecture and debugging, while Hydrogen handles code changes, tests and parallel subagents. In the Antigravity environment, the Orchestrator can switch between models within the same session.

### Is Gemini 4 Hydrogen available on the free tier of the API?

Yes. Hydrogen is available on both the free and the paid tier of the Gemini API, like Gemini 3.8 Flash and Gemini 3.5 Flash-Lite. Carbon and Argon are not available on the free tier. Note, however: to offer applications built on the Gemini API to users in the European Economic Area, Switzerland or the United Kingdom, Google's terms allow only paid services.

### Which Gemini app plans include Hydrogen, Argon and Carbon?

According to Google's guide consulted on 14 November 2026, Google AI Plus includes Hydrogen, Google AI Pro includes Hydrogen and Argon, and Google AI Ultra includes Hydrogen, Argon and Carbon. This applies to personal Google accounts; with a work or school account, models and features depend on the Google Workspace license. Without a paid plan, automatic selection may use Hydrogen for prompts that need more reasoning. Google warns that the list may change without notice.

### Do Hydrogen, Argon and Carbon have the same context window?

Yes. All three share a single 1,048,576-token window across input, reasoning and output, with output of up to 1 million tokens through long decode. They differ in the default reasoning level (high for Hydrogen, medium for Argon and Carbon) and in the maximum output per non-streaming step via REST.

### When do the prices of the Gemini 4 models and Gemini 3.8 Flash change?

The introductory price for Carbon and Argon ends on 30 November 2026: from 1 December 2026 they move to $4.00 for input and $20.00 for output. Gemini 3.8 Flash moves to $1.50 and $7.50 on 1 January 2027. Hydrogen's price stays at the list price, flat across the whole window.
