← all posts
// cost · gemini

Gemini 4 Hydrogen, Neon and Carbon: the lineup makes sense only after the December 1 price change

Google put three Gemini 4 models in front of developers between November 10 and November 18. Carbon arrived first, then Hydrogen, then Neon. If you read the rate card today it looks upside down: Carbon, the model Google recommends for the most demanding work, costs less per token than Neon, the model positioned below it. That inversion has an expiry date, and I think the expiry date is the most useful fact in the whole launch.

The rate card, today and from December 1

ModelRoleInputOutputFree API tier
Gemini 4 HydrogenFast everyday coding and agents1.507.50Yes
Gemini 4 NeonCoding, analysis, business workflows3.0015.00No
Gemini 4 Carbon, through Nov 30Hardest long-running projects2.0010.00No
Gemini 4 Carbon, from Dec 1Same model4.0020.00No

Prices are dollars per million tokens. Carbon's $2 and $10 is an introductory rate that Google's announcement says moves to $4 and $20 afterwards, and Argon, which became generally available on October 13, is on the same schedule. Neon also lists cached input at $0.15, a 95% discount, and Carbon's cached tokens get the same 95% off.

All three share a context window of about one million tokens. Neon has four thinking levels from low to extra high. Hydrogen is the only one with a free tier on the API.

A worked month

Take an agent workload of 100 million input tokens and 20 million output tokens in a month, with no caching, which is pessimistic but keeps the arithmetic honest.

ModelInput costOutput costMonth
Hydrogen150150300
Carbon in November200200400
Neon300300600
Carbon from December400400800

That is my arithmetic from the list prices, not a vendor figure. The same workload on Carbon doubles overnight on December 1 without a single line of your code changing. If you pointed anything at Carbon during the introductory window because it was both the strongest and the cheaper option, this is the bill to plan for.

Google reports that Neon costs 32% less per completed task than Carbon at Neon's default thinking level. Per token, Neon is 25% below Carbon's December rate, so a 32% per-task figure implies Neon also uses somewhat fewer tokens per task in Google's testing. I could not confirm from the coverage I read which Carbon rate that comparison uses, so I would not put the 32% in a budget until I had measured it on my own tasks.

An introductory price is a loan against your December budget. Read the end date before you read the benchmark.

Hydrogen is a Flash successor that left the Flash price class

Hydrogen replaces Gemini 3.8 Flash as the fast model, and Google says it completes tasks 2.5 times as fast as 3.8 Flash at its default reasoning setting. It also costs about double what 3.8 Flash costs. VentureBeat's comparison is the one that stuck with me: Hydrogen's uncached rates are 15 times the lowest standard rates of Claude Haiku 5.5, GPT-6 Luna and Meta's Muse Ember, which all start at $0.10 input.

So Hydrogen is a small reasoning model priced like last year's mid tier, and it should be judged that way. Vals AI ranks it sixth of 41 models on its index at 68.49%, which is a strong result for the smallest model in a family. The mistake would be to drop it into a slot that used to hold a Flash model doing classification or extraction at volume. That slot now belongs to the ten-cent tier or to a decision model. I covered the earlier version of this problem in the Gemini 3.8 Flash cost cliff.

What changes in the apps and tools

In the Gemini app for Pro and Ultra subscribers, Neon replaces Argon and Gemini 3.1 Pro. The Antigravity agent in Gemini Managed Agents now uses Hydrogen by default, with Carbon and Argon available as planning orchestrators. If your team uses Antigravity, an out-of-date client does not just miss one model, it misses the whole arrangement.

Carbon also has a history worth remembering. Google announced it on October 28 for a subset of cyber defenders through the Fairwind Program and held it back from developers for thirteen days while it iterated on guardrails. I wrote about the security side of that in the find versus patch gap. A cyber variant of Neon is going to the same vetted group.

How I would budget it

I would put Hydrogen on the default path for interactive coding and short agent loops, and measure cost per finished task against Sol and Haiku 5.5 with effort turned up, because per-token price alone will mislead in both directions. I would use Neon where long documents, video or spreadsheets are in the input and I need the bigger model. And I would treat Carbon as an escalation target with a hard monthly cap that already assumes the December rate.

The limitation here is that Neon is one day old as I write this. The thinking-level defaults, the rate limits and the per-task numbers are all Google's, and the only independent data points so far are a handful of index scores. The rate card is the part I trust, and the rate card says to do the December math now.

#gemini#pricing#cost#google