Gemini 4 Hydrogen API Pricing: What It Costs Compared With Carbon

Google released Gemini 4 Hydrogen on November 12 for coding and agent workflows, with standard Gemini API pricing of $1.50 per million input tokens and $7.50 per million output tokens. It is cheaper per token than Carbon, but the size of that difference changes on December 1.

For a developer choosing an execution model, the useful comparison is the cost of completing the same task with acceptable output. A lower token rate helps; repeated attempts, extra reasoning and longer answers can still erase the saving.

Hydrogen API pricing, including reasoning and caching

The Gemini API pricing page lists these paid rates in US dollars per million tokens. Output pricing includes thinking tokens.

Processing option	Input	Output, including thinking	Implicit cached input
Standard	$1.50	$7.50	$0.075
Batch	$0.75	$3.75	$0.0375
Flex	$0.75	$3.75	$0.0375
Priority	$2.70	$13.50	$0.135

Hydrogen supports implicit caching, not explicit caching. Standard access has a free tier; the pricing page marks free-tier content as usable to improve Google's products and paid-tier content as not used for that purpose. Grounding and other billable tools can add costs, so the token subtotal is not necessarily the invoice total.

Do not build a cost forecast around a cache hit you have not observed. Start with uncached input, then replace that assumption with measured cache usage after running representative requests. A long shared instruction block may offer an opportunity to save money, but a spreadsheet should not silently assume every request receives the cached rate.

Hydrogen versus Carbon: November and December are different

Carbon's standard rate is $2 input and $10 output through November 30. From December 1, its listed rate becomes $4 and $20 respectively. Hydrogen's $1.50/$7.50 rate therefore represents a 25% reduction against the current Carbon token rates, or 62.5% against the December rates, for equal token counts.

Here is a worked example using invented usage, not measured model performance. Suppose a batch of interactive jobs consumes 10 million uncached input tokens and 2 million billable output tokens, including reasoning. Ignore tools and taxes for this calculation.

Standard-rate model	Input subtotal	Output subtotal	Token total
Hydrogen	$15	$15	$30
Carbon through November 30	$20	$20	$40
Carbon from December 1	$40	$40	$80

The difference between $30 and $40 is modest enough that one extra attempt can matter. If Hydrogen needed twice the example's tokens to finish the same work, its $60 subtotal would exceed November Carbon's $40. It would still be below December Carbon's $80. The date and the retry rate both belong in your comparison.

This is why moving all traffic after reading a cheaper price is a poor migration plan. Route a representative slice first, particularly requests where you can verify correctness automatically, and keep the harder cases separate.

Context size does not tell you how much reasoning to buy

Google's launch announcement describes a one-million-token context window with flat input pricing across that window and generation of up to one million output tokens. It also introduces four thinking levels: low, medium, high by default, and xhigh.

Those are capacity and configuration options, not a recommendation to use the maximum. For a function with a clear specification, extra reasoning might produce no useful improvement. For a change involving ambiguous requirements or interacting components, the same additional work may prevent an expensive mistake.

I would compare a small set of repeated tasks at more than one thinking level: one bounded code edit, one failure diagnosis and one document-to-structured-output job. Hold the instructions and acceptance checks fixed. Record time, billable tokens, correct completions and manual repairs. This is an evaluation plan, not a report of testing Hydrogen.

Pick a stopping rule before running it. For example, reject a cheaper configuration if it breaks an existing test, omits a required field or makes a claim unsupported by the supplied document. Otherwise it is easy to count incomplete work as a saving.

API access and the Gemini app are not the same entitlement

The launch covers AI Studio, the Gemini API, enterprise products and consumer surfaces, but their access rules differ. In the Gemini app, Google's footnote says free users receive Hydrogen through automatic model selection when appropriate for the prompt. That is not a promise that every free account gets a manual Hydrogen selector.

Antigravity is another separate product surface, with its own plan and rollout rules. An API rate does not tell you how much subscription quota an Antigravity task will consume, and buying a consumer plan should not be treated as purchasing API tokens.

The strongest reason to try Hydrogen is to stop paying a frontier rate for work that a faster execution model can reliably finish. The word “reliably” is doing real work there: measure the cost of an accepted result, not the cost of the first response.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain model economics and product decisions in content that developers and founders can actually use.

Book a 30-minute call.