On this page · 5 sections ▾
A model tier is a size and price class within one AI family, and a thinking level is a setting that controls how much reasoning the model does before it answers. Bigger tiers and higher thinking levels usually give better results on hard problems, but they take longer and cost more. Google's new Gemini 4 Neon, which 9to5Google reported sits between the Hydrogen and Carbon models, is a clear case of both ideas in one release.
Why do AI families have several tiers?
Running a large model is expensive. Every answer requires a lot of computation, and the largest models need the most. Smaller models answer faster and cost less, but they make more mistakes on complicated work.
So companies offer a range. In the Gemini 4 family, as 9to5Google laid it out, Hydrogen handles everyday help, Carbon is built for the most complex work, and Neon takes the middle slot for deeper analysis. The Gemini app's model picker now labels them by job: "Everyday help," "Deeper analysis" and "Code & complex work."
The idea is the same as choosing a vehicle. A small car is fine for errands, and a truck is better for a heavy load. Tiers let people pay for only as much capability as the task needs.
What is a thinking level?
Many current models reason before they reply. They produce intermediate steps, which are not always shown to the user, and then give a final answer. More steps usually mean a better answer on math, code or multi-part questions, and also more computation.
A thinking level is the dial for that. Neon offers four settings: Low, Medium, High and Extra High, according to the coverage. High is the default. As FoneArena described it, the Extra High setting is meant for demanding engineering, analytical and scientific tasks, while High balances performance and cost.
The practical effect is that one model can act like several. A quick question can run at a low setting, and a tricky one can get more reasoning.
How do tokens and context windows fit in?
AI models read and write in tokens, which are small chunks of text. Pricing is usually quoted per million tokens, with separate rates for input (what you send) and output (what the model writes).




