Google’s model lineup now has a middle. On Nov. 18 the company put Gemini 4 Neon on sale, slotting it between Hydrogen at the top and Carbon below, and priced it at $3 per million input tokens and $15 per million output tokens.
The positioning is unusual in a market where most releases land at either end. Neon is aimed at work that is too demanding for the budget tier and too frequent to run on the flagship: coding that spans several files, knowledge work that moves between documents, and research tasks that take hours rather than seconds.
What Google is selling is context. Neon takes a million tokens in a single pass, enough to hold the spreadsheets, PDFs, charts and recorded meetings that make up an ordinary corporate decision, and returns a structured deliverable rather than a chat reply. The company says it does not charge extra for the length of that window, which is where the arithmetic starts to matter.
Cached input is priced at 15 cents per million tokens. A team that runs the same long document through the model repeatedly, with a short question attached each time, pays the discount on the bulk of the prompt and full price only on the new material. For anyone processing contracts, filings or support transcripts at volume, that discount decides whether the workflow is economic.
The company put a number on the savings. Google says Neon completes a given task at 32% lower cost than Carbon under default settings, a claim that depends on the task and on which of the four thinking levels the user picks. The default sits at the high setting, which spends more tokens to work through a problem before answering. Cheaper settings exist; they have to be chosen.
Only paid tiers get access. There is no free quota, which puts Neon in the part of the catalog that enterprises buy through committed contracts rather than the part developers try on a weekend.
The price list is about to change under it. On Dec. 1, Carbon and Argon move from $2 input and $10 output to $4 and $20, doubling. Neon, at $3 and $15, will then sit 25% below Carbon rather than above it. Read together, the two moves say something about how Google expects the tiers to be used: the old budget model becomes a premium product for buyers who need its particular behavior, and the new middle takes over the volume.
Doubling a price is a decision companies make when capacity is short, when a model is being repositioned, or when the cheap tier was mispriced from the start. Google has not said which. What is visible is the shape of the resulting ladder: Argon at the bottom, Carbon above it at twice the old rate, Neon between Carbon and Hydrogen, and Hydrogen alone at the top.
That shape raises a practical question for buyers who standardize on one model. A team that built a workflow on Carbon at $2 now faces a choice between paying twice as much for the same behavior or moving to Neon at $3, which is cheaper than the new Carbon price but more expensive than the old one. Either way the unit cost of the workflow rises, and the migration itself costs engineering time.
Long-context work has been the hardest part of this market to price. Providers have generally charged more for bigger windows because attention over a long sequence is expensive to compute. Offering a million tokens at a flat rate is a bet that the underlying serving stack has become efficient enough to absorb the cost, and that customers will pay for the convenience of not splitting documents into pieces.
There is a second bet buried in the pricing. Structured output is what enterprises actually want: a filled table, a reconciled schedule, a summary with citations. Getting a model to produce that reliably is a different problem from getting it to answer questions well, and the vendors that solve it can charge for completed work rather than for tokens consumed. Neon’s framing, with its emphasis on deliverables and its default to a high reasoning setting, points in that direction.
Naming is the part buyers will complain about. Hydrogen, Carbon, Neon and Argon describe no ordering that a customer can guess, and the numbers in the version string do not track the price. Enterprises that run several models behind one interface have to maintain a mapping table, and every release forces a decision about which workloads to move.
The competitive pressure runs both ways. Rivals have been converging on a three-tier catalog for the same reason Google has: a cheap model for classification and extraction, a workhorse for everyday production, and a frontier model for the tasks that justify a large bill. The workhorse slot is where volume lives, and it is the slot every lab is now fighting over.
Whether Neon wins that slot will be settled by benchmarks that buyers run themselves, on their own documents, at their own token counts. The 32% figure is Google’s, measured under default settings on tasks Google chose. Independent testing tends to widen the gaps on some workloads and erase them on others.
For now the catalog has a new middle rung, a date when the rungs below it get more expensive, and a discount on repeated input that rewards the specific behavior Google wants customers to adopt. The decision in front of most engineering teams this month is not whether Neon is better, but whether switching costs less than the price rise already scheduled.


