Gemini 4's three models tested: The most expensive Carbon only wins half the battle
Google released three Gemini 4 models at once. Artificial Analysis' hands-on tests found: the priciest Carbon leads only at low reasoning; once Neon and Hydrogen increase reasoning, their cost-effectiveness overtakes it. Which one to choose? The data decides.
Google released Gemini 4 as three versions at once: Carbon, Neon, and Hydrogen. Artificial Analysis' test data is already out, and the conclusion is counterintuitive: the most expensive Carbon only wins half the battle.
How much does the pricing differ?
| Model | Input price ($/million tokens) | Output price ($/million tokens) |
|---|---|---|
| Carbon | $4 | $20 |
| Neon | $3 | $15 |
| Hydrogen | $1.50 | $7.50 |
Low reasoning: Carbon is strongest, but the cost is tokens
In low-reasoning mode, Carbon performs best among the three, but it also consumes more tokens for the same task. Neon and Hydrogen scale more efficiently at higher reasoning levels, and they also have an extra xhigh option.
Carbon's weak spot
Carbon took first place on CritPt (graduate-level physics evaluation), but lost in agentic research and knowledge work. The reason is that its improvement from low to high reasoning is clearly smaller than that of Neon and Hydrogen.
The Pareto frontier is fully covered by Neon and Hydrogen
Neon and Hydrogen occupy the entire Pareto frontier for intelligence vs cost per task and intelligence vs time per task. Together, the two siblings range from Hydrogen at low reasoning (intelligence index 27, $0.29 per task) to Neon at xhigh (intelligence index 61, $6.44 per task); developers can finely trade off cost and intelligence by adjusting the reasoning level.
Carbon is not on this trend line. Across its reasoning levels, the differences in time and cost are very small.
Someone has already crunched the numbers
Cameron Peterson crunched the numbers: Neon's default high mode has an intelligence index of 59.6, almost matching Carbon's 60.0, while costing only $3.74 per task versus $7.76 for Carbon. Switching to xhigh reaches 61.4, exactly the number Google promoted at launch. His conclusion is blunt: Neon matches Carbon in high mode at half the cost.
Other users are weighing in too. Some mention that although Hydrogen is cheap, it has the fewest hallucinations and may actually be more reliable on many tasks. Some feel Carbon is nearly three times as expensive as Hydrogen, but in daily use the gap is imperceptible. Others point out that Neon and Hydrogen both offer xhigh, while Carbon only goes up to high; choosing a model based on reasoning effort is a good approach. An even more direct suggestion: don't rush to conclusions—run your actual workload first, then decide.
How to choose
If your tasks are not complex, Carbon's low-reasoning mode is indeed strong. But if you are willing to raise the reasoning level, Neon and Hydrogen show a clearer cost-effectiveness advantage.
Full comparison data: https://artificialanalysis.ai/models/comparisons/gemini-4-neon-vs-gemini-4-carbon
发布时间: 2026-11-20 05:39