Gemini 4 Hydrogen vs 3.8 Flash: input and output prices match Jan. 1

Google released Gemini 4 Hydrogen on November 12, 2026, and names it the recommended upgrade from Gemini 3.8 Flash. On the paid tier, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. That's twice 3.8 Flash's current rate, but only until December 31: on January 1, 2027, 3.8 Flash rises to the same $1.50 and $7.50. For teams running Flash in production, the decision is when to move traffic and which settings to change first.

The price is not the only change. Hydrogen defaults to the high thinking level where 3.8 Flash defaults to medium, and it drops explicit context caching. A straight model-ID swap can raise output spend through extra thinking tokens and break requests that still send cached_content, even after the input and output prices match.

What changed with Gemini 4 Hydrogen?

Hydrogen is the fast execution model in the Gemini 4 family. Google pitches it for everyday coding and agent work, alongside its frontier Gemini 4 Carbon and Argon models. Google's developer guide lists it as gemini-4-hydrogen, available on the free and paid tiers of the Gemini API and in Google AI Studio, through both GenerateContent and the Interactions API.

• Context: one 1,048,576-token window shared by input, thinking and output, with long decode up to 1M tokens.
• Thinking levels: low, medium, high (default) and a new xhigh.
• Price: flat across the full window, with no long-context surcharge.
• Where else it runs: Gemini for business, the Gemini app, Google Search, and Google Antigravity for paid AI Pro and Ultra subscribers. Google's guide also says the Antigravity agent in Gemini Managed Agents now uses Hydrogen by default.

Google says Hydrogen finishes tasks 2.5 times faster than 3.8 Flash at its default setting, scores 57% on Terminal-Bench 4.1 against 18.2% for 3.8 Flash, and cuts fabricated claims by more than 5x. Those are Google's figures, not independent measurements.

How much does Gemini 4 Hydrogen cost compared with 3.8 Flash?

Until December 31, 2026, Hydrogen costs twice as much per token as 3.8 Flash. From January 1, 2027, the input and output rates are identical. The table comes from Google's Gemini API pricing page, paid tier, per million tokens.

Scroll sideways to see all 4 columns.

Comparison dimension	3.8 Flash through Dec. 31, 2026	3.8 Flash from Jan. 1, 2027	4 Hydrogen
Input	$0.75	$1.50	$1.50
Output, including thinking	$3.75	$7.50	$7.50
Cached input	$0.075	$0.15	$0.075
Explicit caching	Yes, plus $0.50 per 1M tokens per hour of storage	Yes, plus $1.00 per 1M tokens per hour	No, implicit only
Default thinking level	medium	medium	high

Here's an illustration at list prices. Take a workload of 100,000 requests a month, each with 3,000 input tokens and 800 output tokens, thinking included.

Scroll sideways to see all 4 columns.

100,000 requests a month	Input (300M tokens)	Output (80M tokens)	Total
3.8 Flash, through Dec. 31	$225	$300	$525
4 Hydrogen	$450	$600	$1,050
3.8 Flash, from Jan. 1	$450	$600	$1,050

That's BaristaLabs arithmetic, not a Google figure. It assumes the same token counts on both models, which may not hold: thinking tokens are billed as output, and Hydrogen's high default can think more per request than Flash's medium. Caching and grounding charges are left out.

One cost line moves the other way. Hydrogen's cached input is $0.075 per million with no storage fee, while 3.8 Flash's cached rate doubles to $0.15 in January and its explicit caches bill storage by the hour. Workloads that reuse long prompts may come out ahead on Hydrogen once both models are at full price.

What breaks when you switch from 3.8 Flash to Hydrogen?

The model ID is the obvious change. Google's migration checklist lists several more, and four of them affect cost or can break requests.

Explicit caching is gone. Hydrogen "does not support explicit caching (cachedContents and caches.create)." Requests that reference a cache need the cached_content field removed. Implicit caching runs by default on prompts of at least 4,096 tokens and discounts cache hits by 95%. Put the long, stable part of the prompt first so repeated requests share a prefix.

Thinking defaults to high. If you never set thinking_level, your requests move from medium to high thinking. Google's guide says to "test medium and high on your evaluation suite." Set the level explicitly rather than inheriting the new default. Hydrogen rejects minimal with an error; 3.8 Flash didn't accept it either, but Gemini 3.6 Flash did, so teams jumping from 3.6 should check for it.

Low output caps can cut answers off. Google's thinking guide says max_output_tokens caps thinking and output together. If the model reaches the cap while reasoning, the response comes back "incomplete" with truncated or empty output, "while still billing for any thinking tokens generated." The migration guide tells apps with hardcoded caps "such as 8,192" to raise them to 65,536 and to lower the thinking level instead when they need fewer tokens.

Sampling settings are ignored. Hydrogen ignores custom temperature, top_p and top_k values. Our post on the three settings Gemini 3.6 Flash ignores covers the same behavior on the previous Flash model. If your app relies on low temperature for repeatable output, it has to get that consistency some other way, for example by using structured outputs to fix the response format.

Two smaller items are on Google's list as well. Prompting the model to write raw-text JSON tool calls with no declared tools "can return MALFORMED_FUNCTION_CALL," so declare tools natively or use structured outputs. Multi-turn sessions should pass previous_interaction_id or return thought signatures unchanged.

Who should move before January 1?

BaristaLabs' interpretation: the January price change makes the timing decision for most teams, but it doesn't make the testing optional.

• Agentic and coding workloads on 3.8 Flash. These are the tasks Google pitches Hydrogen for. If your evaluations show fewer retries or failed tool calls, a higher per-token rate in November and December can still cost less per finished task.
• High-volume, simple calls. Classification, routing and short extraction get little from deeper reasoning. Google's own Hydrogen guide says to keep Gemini 3.5 Flash-Lite "for high-volume classification, routing, and sub-second latency targets." At $0.30 input and $2.50 output, it stays cheaper than either Flash or Hydrogen after January 1.
• Apps built on explicit caching. Test the move now. The request changes are mechanical, and implicit caching only pays off if repeated requests share a prefix.
• Teams still on Gemini 2.5 Pro. Google's pricing page says that from December 1, 2026, gemini-2.5-pro requests "will automatically route to Gemini 3.8 Flash" at 3.8 Flash rates. Those teams get a forced move a month before the Flash price doubles, so testing Hydrogen at the same time can save a second migration.

Teams calling Hydrogen through Vercel AI Gateway use the model name google/gemini-4-hydrogen. Vercel's changelog describes Hydrogen as the successor to 3.8 Flash "in the same latency class."

Why it matters for teams paying for Gemini usage

The cheap 3.8 Flash rates end December 31, about seven weeks away. From January 1, every team on Flash pays Hydrogen's input and output rates whether it migrates or not. After that, much of the bill depends on settings: how many thinking tokens each request uses, whether long prompts hit the implicit cache, and whether output caps waste billed thinking.

Data terms are unchanged. Content on the free tier is "used to improve our products," and paid-tier content is not, as with earlier models. Keep customer data out of free-tier experiments, the same project-level care as when Gemini API began storing requests by default.

If your team also uses Gemini 4 Carbon, Google's guide says that adding or switching to Hydrogen from Carbon or Argon "requires no request schema changes." Our post on Gemini 4 Carbon in GitHub Copilot covers Carbon's own price change on December 1.

Recommended first step

BaristaLabs recommends running one production Flash workload on Hydrogen in a paid-tier test project before December, with every changed setting made explicit.

1. Search your code for cached_content, caches.create, thinking_level, max_output_tokens, temperature, top_p and top_k.
2. Pick 200 to 500 recent requests with known good outputs.
3. Run them on 3.8 Flash at its current settings, then on Hydrogen at medium and at high, with max_output_tokens at 65,536.
4. Record quality, latency, and input, output and thought tokens per request from each response's usage fields.
5. Price a month at your real volume at January rates for both models, and switch the workloads where Hydrogen wins on cost per acceptable answer.

For help scoping that test, see BaristaLabs AI consulting.

Sources and scope

• Google, Introducing Gemini 4 Hydrogen, November 12, 2026.
• Google AI Studio, What's new in Gemini 4 Hydrogen, Gemini API pricing and Gemini thinking, accessed November 13, 2026.
• Vercel, Gemini 4 Hydrogen now available on AI Gateway, November 12, 2026.

BaristaLabs did not run Hydrogen or 3.8 Flash for this post. Speed and benchmark figures are Google's, and the cost table is illustrative arithmetic at list prices.