Michael H

Posted on Nov 13 Originally published at agent5.news

Gemini 4 Hydrogen Explained: Speed, Benchmarks, and Costs

#ai #machinelearning #gemini4 #googledeepmind

Originally published on agent5.news.

Google spent most of October 2026 releasing its heaviest new models. Gemini 4 Argon arrived for developers on October 13, followed by Gemini 4 Carbon just two weeks later. Both were billed as frontier-class systems for the hardest problems anyone would throw at an AI. Then, on November 12, Google filled in the third tier: Gemini 4 Hydrogen, a fast, broad-access model built for the everyday tasks that make up the bulk of real-world AI usage. If you have been trying to understand why Google now has a periodic-table naming scheme and what any of it means for your work or your budget, this explainer is for you.

What the Gemini 4 Family Actually Looks Like

Google's Gemini 4 generation now contains three distinct models, each named after a chemical element. Gemini 4 Argon was the first in the family, announced on September 30 for Fairwind cyber defenders and opening to paying developers on October 13. It was described by Google as a frontier model built for deep, long-horizon reasoning across software engineering, enterprise knowledge work, and cybersecurity defense. Gemini 4 Carbon was announced on October 28 and is positioned by Google as its most capable model for sustained long-horizon tasks. Gemini 4 Hydrogen, launched November 12, sits at a different point in the lineup: not a weaker model so much as a differently optimized one, built for rapid execution across a wide range of coding and agentic tasks rather than for the slowest, deepest reasoning passes.

Google's own framing for the division of labor is fairly clear. Carbon and Argon handle deep, long-horizon reasoning on the hardest problems, while Hydrogen serves as a fast execution engine, suitable for everyday coding, agent workflows, and multi-step data processing. That does not mean Hydrogen is a throwaway tier. As the benchmark section below shows, it actually outscores Argon on at least one independent intelligence index.

Why Google Retired the Flash Brand

In Google's earlier Gemini generations, the fast, low-cost tier carried the word Flash: Gemini 2.5 Flash, Gemini 3.8 Flash, and so on. Gemini 4 Hydrogen is explicitly the successor to Gemini 3.8 Flash, which was released on September 2, 2026. Rather than continuing the Flash naming convention, Google moved to the periodic-table scheme that now unifies the entire Gemini 4 family under one coherent brand.

The practical effect for users is that Hydrogen inherits the speed-and-efficiency role that Flash played, but with a new generation of capabilities underneath. Google says the transition is real: according to Google's launch materials, Hydrogen is 2.5 times faster than Gemini 3.8 Flash at the same task at its default thinking setting.

How Hydrogen Performs on Benchmarks

Google published several figures at launch, and at least one comes from an independent evaluator rather than from Google itself.

On Terminal-Bench 4.1, a coding benchmark, Google's launch materials report Hydrogen scored 57 percent compared with 18.2 percent for Gemini 3.8 Flash. That is more than a tripling of the score on the same test.

On the Artificial Analysis Intelligence Index, an independent benchmark maintained by evaluator Artificial Analysis, Hydrogen scores 51 at its default high thinking level. That is one point above Gemini 4 Argon, which scores 50. When Hydrogen is switched to its new xhigh thinking tier, the score rises to 54. For comparison, Gemini 4 Carbon scores around 60 on the same index, based on available sources.

Google also says fabricated claims fell by more than fivefold compared with Gemini 3.8 Flash, though no independent verification of that figure was available at the time of writing.

In Google Sheets, Google says Hydrogen generates outputs faster than previous models in Sheets canvas, which matters for the enterprise workflows the company is targeting with that integration.

One third-party evaluation from a biotechnology company called LatchBio tested Hydrogen on an antibody development benchmark, suggesting the model's factual reliability holds up in specialized scientific contexts, though the specific figures have not been independently verified.

The Four Thinking Levels: What They Mean

One of Hydrogen's most practically significant features is its system of four tunable thinking levels: low, medium, high (the default), and a new tier called xhigh. Hydrogen extends all the way to xhigh, offering more granularity than Carbon or Argon.

The thinking levels let developers and users trade off reasoning depth, response latency, and cost. At the low end, tasks run fast and cheaply. At xhigh, the model reasons more extensively but consumes more output tokens compared with the high setting, which raises the cost of a typical query.

Google's pitch here is that you should not pay for reasoning depth you do not need. A simple code completion does not require the same thinking budget as debugging a complex multi-file refactor, and Hydrogen's levels let you match the tool to the task.

Orchestrator: Carbon and Hydrogen Working Together

The most architecturally interesting aspect of the Hydrogen launch is a new feature called Orchestrator, available inside Google's Antigravity development platform for Google AI Ultra subscribers. Rather than assigning an entire task to a single model, Orchestrator dynamically switches between Carbon and Hydrogen within the same session based on what each step requires.

In a typical Orchestrator workflow, Carbon handles planning and sets the direction for a complex task, then hands off routine execution steps, such as writing edits, updating packages, and running tests, to Hydrogen. If Hydrogen encounters something it cannot resolve, it hands back to Carbon to inspect the state and repair the plan before resuming.

The efficiency gains Google reports from this approach are substantial. According to Google, on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while consuming 24 percent less quota and completing tasks 19 percent faster on average. If those numbers hold in real-world use, Orchestrator could meaningfully change the economics of running complex multi-step AI agents, particularly once Carbon's standard pricing takes effect in December.

What Hydrogen Costs

Pricing is one of Hydrogen's clearest selling points relative to the two larger models.

For developers, Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across its full 1 million-token context window. That pricing applies regardless of which thinking level is selected, so the cost of moving from high to xhigh comes from the additional output tokens the model generates, not from a separate tier surcharge.

For comparison, Gemini 4 Argon currently costs $2.00 per million input tokens and $10.00 per million output tokens during its introductory period, which runs through November 30, 2026, after which it rises to $4.00 and $20.00. Gemini 4 Carbon is on the same introductory rate of $2.00 and $10.00 through the same date, also scheduled to double on December 1.

In other words, Hydrogen is priced at 25 percent less per input token and 25 percent less per output token than Carbon and Argon at their current introductory rates, and 62.5 percent less than those models once their standard December pricing takes effect.

For consumers, access is even simpler. Hydrogen is free for all users in the Gemini app and Google Search through automatic model selection. Free Gemini users get it when Google determines it is the right model for a given prompt. Google AI Pro and Ultra subscribers can access it directly in the Gemini app model picker and in Antigravity.

Where Hydrogen Is Available

Google made Hydrogen available globally the same day it was announced, which is a broader immediate rollout than either Carbon or Argon received at launch. Both of those models started with limited access through the Fairwind Program for cyber defenders before expanding to paying developers.

Hydrogen is available across Google AI Studio and the Gemini API for all tiers, including free tiers. It is in Android Studio for mobile developers, in Gemini for business, in the Gemini app, and in Google Search globally. Paid Google AI Pro and Ultra subscribers can also use it inside Antigravity, including the new Orchestrator feature. Gemini in Google Sheets access starts on Pro and Ultra plans as well.

Hydrogen is active in Google Search globally, which means that many people who use Google Search are already interacting with the model even if they have never heard the name.

What Happens Next

The most concrete near-term event to watch is December 1, 2026, when the introductory pricing for both Gemini 4 Argon and Gemini 4 Carbon doubles. Argon moves from $2.00 and $10.00 to $4.00 and $20.00. Carbon does the same. Hydrogen's pricing has not been announced to follow the same schedule, so the cost gap between the fast tier and the frontier tiers could widen significantly starting next month.

The Orchestrator feature is currently limited to Google AI Ultra subscribers in Antigravity. Google has not announced a timeline for broader availability, so that is worth monitoring if you are on a lower subscription tier or working purely through the API.

Google also noted that Hydrogen includes updates from the Gemini 4 safety and security stack, with built-in training and real-time surface safeguards. As with the safety improvements Google has described for Carbon and Argon, independent red-team evaluations of these claims have not yet been published, so scrutiny from the research community will be an ongoing story.

Finally, there is the question of how the competitive landscape responds. Hydrogen's Artificial Analysis Intelligence Index score of 54 at xhigh is in a similar range to comparable models from other labs at their maximum reasoning settings, though the gap varies by model. Whether that gap narrows or widens with the next round of releases is worth watching closely.

DEV Community