Gemini 4 Hydrogen: Google's new AI model aims to offer near-top performance at a low price

AI at a fraction of the cost: Why Gemini 4 Hydrogen is putting pressure on the market

Pay less, get almost the same performance: Google's new strategy with Gemini 4 Hydrogen

With Gemini 4 Hydrogen, Google is positioning an affordable AI model that promises significantly more performance than a typical all-rounder: In several independent comparisons, it achieves near-peak performance levels in programming tasks, knowledge work, and agent-based processes, often surpassing significantly more expensive top-of-the-line models – and in some cases even exceeding Google's own, higher-positioned Gemini 4 Argon. At the same time, Google employs four levels of intelligence, allowing users to tailor computational effort, response quality, and costs to the specific task. For companies and developers, this means that not only the performance of a model becomes crucial, but also the question of which level of intelligence is economically viable. A look at prices, benchmarks, and application areas reveals where Hydrogen excels, when the premium for more in-depth intelligence is worthwhile, and how Google's new model could reshape the pricing architecture of the AI ​​market.

Almost top quality at a discount price – anyone still paying premium is paying for a prestige that soon nobody will need

When the small model overtakes the large one

There are product announcements that primarily document technological advancements, and those that shift the market landscape. The introduction of the Gemini 4 Hydrogen clearly belongs to the latter category. Google officially positions the model as a fast and cost-effective all-rounder for everyday tasks, programming work, and agent-based workflows. It is intended as the successor to the Gemini 3.8 Flash, the model class that has thus far stood for high speed, low cost, and solid, but not outstanding, performance. However, the Hydrogen transcends this role. In several comparisons, it outperforms the larger and ostensibly higher-end Gemini 4 Argon.

The economic irony lies in the timing. Just two days before Hydrogen, Google unveiled its Frontier model, Gemini 4 Carbon, and on that occasion recommended using Argon for everyday programming. With Hydrogen, this recommendation is effectively obsolete. Argon, which has only been on the market for barely four weeks, thus loses its economic justification in large parts of its application spectrum. A product that goes from recommendation to footnote in less than a month says a lot about the pace and logic of the current AI market.

For companies, developers, and investors, this is more than just a footnote. It shows that the performance boundary between premium and mass-market models is blurring faster than price lists and procurement processes can reflect. Anyone planning AI budgets today must expect a half-life of model decisions measured in weeks rather than years.

The price structure as a strategic signal

Google charges $1.50 per million input tokens and $7.50 per million output tokens for Hydrogen. Cached hits, meaning previously processed and reused contextual content, cost only $0.075 per million tokens. The context window can hold up to one million input tokens, plus up to one million output tokens for multimodal knowledge work.

Behind these figures lies a well-thought-out pricing logic. The ratio of output to input cost is five to one. This reflects the actual computing costs, as generating text is sequential and more computationally intensive than parallel reading. At the same time, this ratio creates an incentive: applications that do a lot of reading and little writing benefit disproportionately. These include document analysis, research, contract review, classification, and summarization—precisely the areas where companies currently locate the majority of their productive AI use.

Even more remarkable is the price for cached hits. It's exactly five percent of the regular input price, which equates to a 95 percent discount. Economically, this means that anyone using large, recurring contexts, such as a company-wide rule set, an extensive codebase, or a product database, pays almost nothing for repeated use. A fully populated context window of one million tokens costs $1.50 for the initial loading, and only 7.5 cents for each subsequent use from the cache. This shifts the cost structure from variable costs per request to a kind of fixed-cost model for knowledge that is loaded once and then used many times.

This architecture is no accident. It rewards customer loyalty and deep integration. Once a company has anchored its data in a specific infrastructure and optimized its caching logic, it won't easily switch to a competitor. The low entry price is therefore also a tool for generating switching costs.

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

Levels of thought as second-degree price differentiation

Perhaps the most important economic innovation of Hydrogen is its four levels of intelligence: low, medium, and high as the standard, and the new xhigh level. The higher the level, the more internal calculations the model performs before responding. This improves the quality of the results but also increases the number of tokens generated and therefore the cost.

From a microeconomic perspective, this is a classic example of second-degree price discrimination. The provider does not know the individual customer's willingness to pay, but offers a menu of different quality levels from which each user can choose the appropriate combination of price and performance. A single model thus serves both the mass market with simple routine tasks and demanding users with complex problems.

The values ​​determined by independent analysts make the mechanics transparent. At the xhigh level, an average task costs around $3.16 with an intelligence index of 54 points. At the high level, the cost drops to $1.87 with 51 points. The medium level costs $0.80 and achieves 41 points, while the low level costs only $0.29 with 27 points.

Converting these values ​​into costs per index point reveals a revealing picture. At the low level, an index point costs approximately 1.1 cents, at medium just under 2 cents, at high around 3.7 cents, and at xhigh about 5.9 cents. The average cost of intelligence thus increases significantly with each level.

The picture becomes even clearer when considering marginal costs. The jump from low to medium costs an additional 51 cents and yields 14 index points, or roughly 3.6 cents per additional point. The jump from medium to high costs $1.07 and yields 10 points, which is about 10.7 cents per point. The final step, from high to xhigh, costs $1.29 for just three additional points, which corresponds to roughly 43 cents per point. The marginal cost of the final quality level is therefore about twelve times higher than that of the first quality jump.

This is a textbook example of diminishing marginal returns. At the xhigh level, the model requires almost twice as many output tokens as at the high level, but delivers only a moderate improvement in quality. This leads to a clear recommendation for practical application: The standard high level is the economic sweet spot for most demanding applications. The xhigh level is only worthwhile where a single error is costly, such as in legal assessments, security-relevant code, or decisions with high financial leverage. The low level is suitable for high-volume tasks of low complexity where speed and volume are more important than depth.

Independent analysts also confirm that Hydrogen, at its high, medium, and low settings, falls on the so-called Pareto boundary between processing time and intelligence. This means that currently no other model is both faster and more intelligent. For companies that need to balance latency and quality, this is a compelling argument.

Competitive comparison: just behind the leader, but far ahead in price

In the independent Intelligence Index, Hydrogen achieves 54 points at the highest cognitive level. This puts it just one point behind Claude Sonnet 5.5 and GPT-6.1 Sol, which each score 55 points at maximum cognitive level. At the standard high level, Hydrogen achieves 51 points, placing it just behind GPT-6.1 Sol at the xhigh level and Claude Opus 5.5 at the high level, both of which reach 52 points. The in-house model Argon only manages 50 points, while the Frontier model Carbon achieves 59.

More interesting than the raw scores, however, is the performance-to-cost ratio. At the xhigh level, Hydrogen costs about a third per task of what Claude Sonnet 5.5 costs at maximum level, with nearly identical performance. Roughly speaking, this means that a comparable task costs around $9.50 with Sonnet, but only $3.16 with Hydrogen. At the same time, Hydrogen at this level is about two and a half times more expensive than GPT-6.1 Sol at maximum level, which therefore costs around $1.26 per task and achieves one more point.

This situation demonstrates that Google is by no means unrivaled in its low price. The market has split into three price segments. Anthropic positions itself in the upper price segment, justifying this with strengths in specific disciplines. OpenAI, with its GPT-6.1 Sol at the maximum level, offers a particularly aggressive price-performance ratio. Google occupies the middle ground with Hydrogen, but emphasizes the flexibility of its development stages and its integration into a comprehensive product ecosystem.

A numerical example illustrates the scale of the costs. A company executing 100,000 demanding agent tasks per month would spend approximately $316,000 on Hydrogen at the xhigh level, around $187,000 at the high level, and only $80,000 at the medium level. With Claude Sonnet 5.5 at its maximum level, the costs would be around $950,000, and with GPT-6.1 Sol at its maximum level, approximately $126,000. The difference between the cheapest and most expensive option with comparable peak performance is therefore more than sevenfold. For the economic viability of AI projects, the choice of model is thus no longer a technical detail, but a strategic cost decision of considerable consequence.

Performance improvement compared to its predecessor

Compared to Gemini 3.8 Flash, the progress is considerable. According to Google, Hydrogen completes tasks at the standard level two and a half times faster. In Terminal Bench 4.1, a test for agent-based command-line work, the score increases from 18.2 to 57 percent, more than threefold. The rate of fabricated statements is said to have decreased by more than five times.

For economic efficiency, these three factors are closely linked. Speed ​​not only reduces waiting times for users but also the cost per completed step in automated process chains. A higher success rate reduces the number of repetitions, and every avoided repetition saves tokens and time. Finally, fewer hallucinations lower the costs of human oversight, which is the real bottleneck in many companies when using generative AI.

An often underestimated aspect is the improved handling of failed tool calls and unsuccessful search queries. In agent-based systems, where a model independently calls tools, queries databases, or searches websites, errors are unavoidable. The crucial factor is how the model reacts to them. An agent that gets stuck in a loop after an error or simply invents missing information incurs follow-up costs that quickly negate any price advantage. An agent that recognizes errors, seeks alternative paths, or transparently reports that it cannot proceed is far more valuable in production environments. This robustness is difficult to capture in a single benchmark figure, but in practice, it determines the success of automation projects.

Knowledge work as the real strength

Hydrogen performs particularly well in knowledge work. In the Vals Tax Agent Bench, which tests tax-related tasks, the model achieves 52.3 percent, securing first place. In the Vals Legal Research Bench for legal research, it shares first place with 55.8 percent. For open-ended agent tasks such as research, data analysis, and code navigation, Hydrogen ranks fifth in the Agent Arena. It achieves 71.6 percent in the Agents' Last Exam test, 76.1 percent in the AutomationBench, and 53.4 percent in the multimodal knowledge agent test. In Vending Bench 2, a simulation in which an agent must operate a vending machine profitably over an extended period, Hydrogen achieves a balance of $14,549, surpassing Argon's $13,718, though falling short of Carbon's

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs