[KongchangAI]
· 3 min read· 1,507 words

Gemini 4 Hydrogen Released: Low-Cost Model Pricing Strategy and Performance Comparison

Gemini 4 Hydrogen Released: Low-Cost Model Pricing Strategy and Performance Comparison

Gemini 4 Hydrogen completes the product line, but with limited discounts and noticeably reduced capabilities, it offers insufficient value.

Google released Hydrogen, the cheapest Gemini 4 model, but its pricing is confusing: at $1.50/M input and $7.50/M output, it is only 25% cheaper than Carbon/Argon, with the discount reaching 75% only after Carbon/Argon prices double on December 1, 2026. Carbon and Argon cost the same, but Argon is faster and uses fewer tokens. Pelican SVG tests show Hydrogen produces far less detail but is faster. Google positions it as a daily execution engine and new free-tier model. Yet at ~15x the price of Claude Haiku 5.5 and GPT-6 Luna ($0.10/$0.50), it lacks short-term price appeal.

Google today released Gemini 4 Hydrogen, the lowest-cost member of the Gemini 4 family. However, its pricing strategy is somewhat confusing—the price gap between Hydrogen and the higher-end Carbon and Argon is not large, yet there are clear differences in performance. Based on the release information for llm-gemini 0.37 and hands-on testing, this article analyzes the pricing logic and model positioning of the Gemini 4 series.

rss source: llm-gemini 0.37

Pricing Structure: Carbon and Argon at the Same Price, Limited Hydrogen Discount

According to the official pricing table, Gemini 4 Carbon and Argon currently cost $2.00/M tokens for input, $0.10/M for cached input, and $10.00/M for output. Both are scheduled to double in price on December 1, 2026, at which point input will be $4.00/M and output $20.00/M. Hydrogen is priced at $1.50/M for input, $0.075/M for cached input, and $7.50/M for output.

This means that before December 1, Hydrogen offers only a 25% discount compared with Carbon/Argon; after December 1, because the Carbon/Argon prices double, Hydrogen's discount will reach 75%. This pricing strategy makes Hydrogen appear to offer insufficient value for money at the current stage.

Even more puzzling is why Carbon and Argon are priced exactly the same yet split into two models. Google's explanation is that Argon provides faster response times and lower token consumption per interaction than Carbon. In other words, although the unit price is the same, Argon reduces actual costs by using fewer tokens.

Cached Input is a discount mechanism in large-model API billing: when the prompt prefix of a request is highly repetitive with recent requests (such as a fixed system prompt, a long document, or conversation history), the provider can mark this portion of tokens as a cache hit and bill it at a price far below standard input. In the original text, Carbon/Argon's standard input is $2.00/M, while cached input is only $0.10/M, equivalent to a 95% discount; for Hydrogen it is $1.50/M versus $0.075/M. The caching mechanism has a huge impact on scenarios that repeatedly call the same long context at high frequency (such as codebase Q&A and customer service bots), and the actual cost may be far below the nominal price. Therefore, when comparing the cost-effectiveness of models, you should not only look at the standard input/output unit prices but also estimate comprehensively based on cache hit rates and the magnitude of cache discounts.

Performance Comparison: Hydrogen's Pelican Is Clearly Simpler

To compare the capabilities of the three models intuitively, I asked each of them to generate an SVG image of "a pelican riding a bicycle," and set three thinking levels: high, medium, and low. The results differed significantly.

At the high thinking level, Carbon generated a pelican wearing a sailor hat and riding a red bicycle with fenders and a wicker basket, with a fish in the basket, against a background of a beach, sun, clouds, and flying birds; Argon's pelican wore a red hat and a flowing teal scarf, with a blue fish in the bicycle basket, against a background of a seaside promenade; while Hydrogen generated a very simple pelican: a white pelican riding a basic red bicycle, with only a pale blue sky, two clouds, speed lines, and green ground in the background.

At the medium and low thinking levels, the trend was similar: Carbon and Argon could generate richly detailed seaside scenes, while Hydrogen's images were much simpler. In terms of generation time, Hydrogen was also noticeably faster.

Despite the small price gap, the gap in image generation quality for Hydrogen is quite obvious. This may reflect Google's positioning of Hydrogen: not the pursuit of extreme capability, but the pursuit of efficiency and cost.

Thinking level refers to the "internal thinking" budget that a reasoning-capable model can invest before generating an answer, usually divided into high, medium, low, or off. At a high thinking level, the model generates more intermediate reasoning tokens before giving the final answer, excelling at complex logic, multi-step planning, and detail-rich generation tasks, but with higher latency and more token consumption; a low thinking level skips or compresses the reasoning process, resulting in faster responses and lower cost, but with reduced completion quality on complex tasks. Different vendors name this parameter differently (such as reasoning effort or thinking budget), but the essence is to let users trade off among quality, latency, and expense. The comparison of the three thinking levels in the article is precisely to observe the differences in the capability ceiling of the same model under different reasoning budgets.

Hydrogen's Positioning: Daily Execution Engine and Free-Tier Model

Google describes Hydrogen's positioning in its official blog as follows: in real software development, teams rely on frontier models such as Carbon and Argon for deep architectural judgment and complex bug investigation, but they still need a fast, disciplined execution model to write code, run tests, and verify changes. Hydrogen can serve both as a standalone model for daily work and, paired with Carbon and Argon, as an efficient execution engine in multi-model workflows.

Interestingly, Hydrogen is also the new free-tier model and will power AI features in Google Search. This means that although Hydrogen compromises on capability, it plays an important role in Google's ecosystem—especially for scenarios requiring large-scale, low-cost inference.

Multi-model workflow refers to orchestrating multiple models with different roles simultaneously within a software or business process, letting each perform its own function. A common pattern is "frontier models plan, cheap models execute": capable models handle architecture design, complex bug localization, and other stages requiring deep judgment, and then the decomposed concrete tasks (writing code, running tests, changing configurations) are handed to fast, low-cost models for batch completion. The benefit of this approach is a significant reduction in overall cost while retaining quality in key stages; the cost is the need for additional task decomposition, result validation, and failure retry mechanisms. Hydrogen is positioned as an "execution engine," which is a typical product of this pattern—its value lies not in being the strongest when used alone, but in being a high-throughput stage in the pipeline.

Price Competitiveness: Still Far More Expensive Than Competitors

Although Hydrogen is the cheapest model in the Gemini 4 family, its price is still on the high side compared with lightweight models from other vendors. Claude Haiku 5.5 and GPT-6 Luna are both priced at $0.10/$0.50 (input/output), and Hydrogen is 15 times their price.

This may reflect the overall higher positioning of the Gemini 4 series, where even the "cheap" version maintains relatively premium pricing. Cost-sensitive users may need to wait for the price adjustment after December 1 or consider models from other vendors.

$/M tokens (dollars per million tokens) is a common industry unit for comparing the cost of large-model APIs. The cost of an API call = number of input tokens × input unit price + number of output tokens × output unit price (cached portions are calculated separately). Because output tokens are usually 4–5 times more expensive than input tokens, and the input-output ratio varies greatly across tasks (for example, long-document summarization is input-heavy, while code generation is output-heavy), simply comparing a single unit price can easily be misleading. The 15x price gap in the article is derived from a direct comparison of input/output unit prices; in actual use, if Hydrogen consumes more tokens or has a lower cache hit rate, the real cost gap may further widen or narrow, so model selection should be evaluated in light of the token structure of the specific task.

Summary and Recommendations

The release of Gemini 4 Hydrogen completes the Gemini 4 family product line, but its pricing strategy lacks appeal in the short term. If you need high-quality multimodal generation capabilities, Carbon and Argon are better choices; if you are pursuing low-cost inference, Hydrogen will be more cost-effective after December 1. For now, Hydrogen is better suited for use as a free-tier model or as an execution engine in multi-model workflows.

For developers, llm-gemini 0.37 already supports these new models, making it easy to test and integrate them via the command line.

One-line Summary

Gemini 4 Hydrogen completes the product line, but with limited discounts and noticeably reduced capabilities at its current price, it offers insufficient value for money.

Paragraph Summary

Google released Hydrogen, the cheapest model in the Gemini 4 family, but the pricing strategy is confusing: currently $1.50/M input and $7.50/M output, only 25% cheaper than Carbon/Argon; the discount will expand to 75% only after Carbon/Argon prices double on December 1, 2026. Carbon and Argon are priced the same, with the difference being that Argon responds faster and consumes fewer tokens per interaction. The author tested the three models with "pelican riding a bicycle" SVGs, and Hydrogen's generated images had noticeably fewer details but were faster. Google positions Hydrogen as a daily execution engine and the new free-tier model, powering AI features in Google Search. However, compared with Claude Haiku 5.5 and GPT-6 Luna at $0.10/$0.50, Hydrogen is still about 15 times more expensive and lacks price appeal in the short term.

#Gemini 4 Hydrogen#Gemini 4 pricing#Gemini 4 Carbon#Gemini 4 Argon#LLM API pricing#multi-model workflow#free-tier model
Share:

Related articles