If you run Gemini 3.8 Flash at volume, its token rate doubles on January 1, 2027, and Google's replacement finished Artificial Analysis's benchmark tasks for about a quarter of the cost. Gemini 4 Hydrogen lists at $1.50 per million input tokens and $7.50 per million output tokens, twice what 3.8 Flash charges today. Artificial Analysis measured Hydrogen at its default high setting at $1.87 per Intelligence Index task against $7.77 for 3.8 Flash, with a score of 51 against 31. Move your high-volume agent and coding workloads in the next six weeks, and budget them on cost per task.
The token rate is the wrong number to budget on here. Flash's cheaper tokens bought longer runs, and the Gemini API pricing page now shows Flash climbing to Hydrogen's exact rate on New Year's Day.
What Google Shipped on November 12
Gemini 4 Hydrogen is Google's new everyday model, the tier 3.8 Flash held, and it went generally available on the Gemini API the day it was announced. The Gemini API changelog lists gemini-4-hydrogen as GA on November 12, 2026 at $1.50 and $7.50 per million tokens on the paid tier, with a free tier as well.
Google DeepMind calls it a "daily driver" for agentic work in its launch post, with Gemini 4 Carbon and Gemini 4 Argon kept for longer reasoning jobs. The pricing is flat across a 1 million token context window. Hydrogen ships to the Gemini API, AI Studio, Android Studio, Gemini for business, Google Antigravity and Google Search, and Google's developer post says it can run alone at "four configurable thinking levels" or alongside Carbon and Argon.
The Hydrogen model card says the model "is adapted from Gemini 4 Carbon," has a March 2026 knowledge cutoff, and returns up to 64K output tokens synchronously, or up to 1M through background long decode. The thinking docs name the four thinking levels as low, medium, high (the default) and a new xhigh.
For enterprise buyers on Google Cloud, the Vertex AI model catalog lists Gemini 4 Hydrogen among its generally available Gemini models, next to Gemini 3.8 Flash. Neither carries a retirement date there yet.
Why a Higher Token Price Can Mean a Lower Bill
Hydrogen costs less per finished task because it uses fewer tokens and less time to get there. Cost per task is the total spend to complete one benchmark task, counting input, output and thinking tokens together. It is the number that turns into your invoice.
Here is what Artificial Analysis recorded for both models at the high setting:
| Metric | Gemini 4 Hydrogen (high) | Gemini 3.8 Flash (high) |
|---|---|---|
| Intelligence Index | 51 | 31 |
| Cost per task | $1.87 | $7.77 |
| Cost to run the full index | $1,768 | $5,764 |
| Output tokens for the index | 110M | 126M |
| Time per task | 8.1 min | 14.6 min |
| List price, input / output per 1M | $1.50 / $7.50 | $0.75 / $3.75 |
The output token counts are close, so they do not explain the gap. At Flash's $3.75 output rate, 126M output tokens come to roughly $470 of its $5,764 index bill. The rest is input and cached traffic, which the comparison page does not break down. A plausible reading is that Flash took more turns per agentic task and re-sent its context each time, but that is an inference, and your own logs will tell you whether it holds for your workload.
Google's own numbers point the same way, with the usual caveat that they are vendor claims. The launch post puts Hydrogen at 57% on Terminal-Bench 4.1 against 18.2% for 3.8 Flash, about 2.5x faster at high thinking, and says partners saw 12% to 23% lower token usage on data-heavy workflows.
The xhigh setting is where the savings go
Turning thinking up to xhigh buys three index points at a much higher price. On its model page, Artificial Analysis scores Hydrogen at xhigh at 54, at $3.16 per task, after generating 304M output tokens to run the index, against a 77.3M median. It ranks the model 78th of 79 on verbosity. The model card's known limitations say the model "may use more tokens to maximize performance, especially at higher thinking levels."
Leave the default at high. Route specific task types to xhigh only after an eval shows the extra points change an outcome you care about.
The January 1 Price Step Removes the Reason to Wait
Staying on 3.8 Flash stops saving money on January 1, 2027. The pricing page shows Flash's standard input rate going from $0.75 to $1.50 and its output from $3.75 to $7.50, with the note "$0.75 through December 31, 2026. $1.50 starting January 1, 2027." Batch, Flex and Priority rates double on the same date, and context caching storage moves from $0.50 to $1.00 per million tokens per hour.
From that date the two models list at the same token price, and the cached input rate tips the other way. Flash's cached input goes from $0.075 to $0.15 per million, while Hydrogen's is listed at $0.075 with no change noted. If your agents lean on prompt caching, Flash becomes the more expensive model per cached token.
Two more dates on the same page and changelog belong in your plan:
- December 1, 2026. The changelog says requests to
gemini-2.5-proroute togemini-3.8-flashand are "billed at Gemini 3.8 Flash rates." Anything you still have pointed at 2.5 Pro lands on Flash for one month, then its price doubles. - December 1, 2026, for the frontier tier. Gemini 4 Carbon and Argon go from $2.00 / $10.00 to $4.00 / $20.00 per million tokens, per the pricing page. If your routing sends hard cases up to Carbon or Argon, that escalation path doubles a month before Flash does.
Hydrogen's row carries no scheduled change. Carbon's $2.00 / $10.00 launch rate came with a step to $4.00 / $20.00, which we covered when Carbon launched. Ask for Hydrogen's rate in writing before you build a 2027 budget on it.
Where Hydrogen Is Not the Obvious Pick
Hydrogen beats 3.8 Flash on cost per task, but it is not the cheapest model at its score. Artificial Analysis shows GPT-6.1 Terra at max effort also scoring 51, at $0.96 per task and $972 for the full index, about half of Hydrogen's $1.87 and $1,768. Hydrogen's answer is speed: 174 output tokens per second against Terra's 68. If you are not already committed to Google, that trade is the one to price, and our GPT-6.1 Terra launch piece covers the other side.
Three other things can break a straight swap:
- Refusals went up. The model card reports unjustified refusals 5.57 percentage points higher than 3.8 Flash and says they are higher "particularly for multimodal prompts." If you run document or image extraction on Flash, test those prompts first.
- Gemini Enterprise has not posted a Hydrogen entry. The launch post names Gemini for business as a surface, but when we read the Gemini Enterprise release notes on November 14, 2026 they covered Carbon (GA November 10, off by default in the web and mobile apps) and Argon, and nothing on Hydrogen's admin defaults or data residency.
- We could not confirm a Vertex AI rate. The catalog lists Hydrogen as GA, but we could not confirm Vertex AI pricing for Hydrogen, or whether Flash's January 1 step applies to Vertex contracts, on November 14, 2026. Get both from your Google Cloud account team before you assume API list prices carry over to committed spend.
What to Do Before January 1
The work is a routing change and an eval, and the deadline is the first invoice of 2027.
This Week:
- Pull 30 days of Gemini usage by model ID and find every caller on
gemini-3.8-flashandgemini-2.5-pro. The 2.5 Pro callers move to Flash on December 1 without a code change. - Pick your three highest-volume Flash workloads and log tokens per completed task, input and output separately. That is your baseline. Without it you cannot tell whether Hydrogen's lower cost per task holds on your traffic.
This Month:
- Run those three workloads on
gemini-4-hydrogenat the high setting against your own eval set. Compare cost per completed task, success rate and wall-clock time. Include your multimodal prompts so the refusal regression shows up before production does. - Ask your Google account team three questions in writing: the Vertex AI rate for Hydrogen, whether Flash's January 1 increase applies to your contract, and whether Hydrogen's $1.50 / $7.50 has an end date.
Before January 1:
- Switch the workloads that passed. Keep Flash only where your eval shows it wins, and re-cost those at the 2027 rate.
- Re-baseline your 2027 model budget on cost per task. If you route hard cases to Carbon or Argon, price that escalation at the December 1 rate.
The Bottom Line
Google's pricing page now schedules list-price doubles for three models, Carbon and Argon on December 1 and 3.8 Flash on January 1, and the per-token comparison misleads on this launch. We saw the same pattern when Claude Sonnet 5.5 cost more than Opus 5.5 at the same score and when Meta's Muse Ember came in at $0.14 a task: the cheaper token did not make the cheaper job. FinOps teams that still budget models on rate cards will read Hydrogen as a 100% increase. Teams that log tokens per completed task will see a cut, on Artificial Analysis's tasks at least.
Hydrogen and Flash list at $1.50 / $7.50 from January 1, and only one of them scored 51. Run the eval this month and switch what passes.
Continue Reading
- Gemini 4 Argon Leads Vals at $15.68 a Task With No GA Date
- Google Opens Gemini 4 Carbon to a Subset of Cyber Defenders
- Fable Cost $18 a Fix on Private Code. Flash Cost $8.
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- Google's Antigravity Lacks SOC 2 and ISO 27001 as Code Assist Ends
- Claude vs GPT vs Gemini: Stop Comparing Per-Token Prices