Gemini 4 Hydrogen costs twice Gemini 3.8 Flash per token, and a quarter as much per task

Google released Gemini 4 Hydrogen on November 12, the third Gemini 4 model after Argon and Carbon and the one Google is putting in front of everyone: free Gemini app users, Google Search and the free tier of the Gemini API. In the API it is gemini-4-hydrogen, priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1M-token window. That is exactly twice what Google charges today for Gemini 3.8 Flash, the model Hydrogen is meant to replace.

The independent numbers point the other way. On Artificial Analysis, which runs every model through the same set of agentic and reasoning evaluations, Hydrogen at its default high setting costs $1.87 per Intelligence Index task. Gemini 3.8 Flash at high costs $7.77. A model with double the per-token price comes out about four times cheaper per task, scores 51 against 31, and finishes in 6.0 minutes per task instead of 14.8.

Per-token price is the wrong number for agents

The explanation is in the token counts Artificial Analysis publishes for each run. To get through the index, 3.8 Flash read 10.2 billion input tokens and wrote 126 million. Hydrogen at high read 5.34 billion and wrote 110 million. In agentic work, almost all of the tokens are input: the context gets re-sent on every turn of the loop, and output is between about 1% and 3.5% of the total for every model in the table below. A model that needs fewer turns, or reads less per turn, wins on the bill even with pricier tokens.

The token counts don't explain the whole fourfold gap at list prices, and Artificial Analysis' pages don't give its cost breakdown in text form, so we can't say how much comes from caching. Hydrogen's implicit cache reads are 95% off ($0.075 per million); 3.8 Flash's are 90% off. What we can say is that the per-token comparison, which is how most people will first read the price, gets the direction wrong.

Google's own claims are consistent with this. Koray Kavukcuoglu's announcement says Hydrogen "natively" batches file-read and search calls in parallel instead of opening files one at a time, and that early partners saw it run 2x to 2.8x faster than 3.8 Flash while using 12% to 23% fewer tokens. Those are partner figures we can't check; the Artificial Analysis measurements are the independent ones.

There's also a date that makes the comparison less theoretical. The 3.8 Flash price on Google's pricing page is introductory: $0.75/$3.75 through December 31, 2026, then $1.50/$7.50 from January 1, 2027, the same as Hydrogen, with cache reads at $0.15 instead of Hydrogen's $0.075. And from December 1, Google will answer gemini-2.5-pro calls with 3.8 Flash, as we reported earlier this month. Anyone who lets that routing happen will, a month later, pay Hydrogen's prices for the model that scores 20 points lower.

Where Hydrogen sits against everyone else

Artificial Analysis figures, retrieved November 12. Cost per task is at each provider's list price; for Carbon and Argon the site uses $4/$20, the rate that applies from December 1 (Carbon is $2/$10 until then).

Three things stand out.

Inside Google's lineup, Hydrogen makes Argon hard to justify. At high it edges Argon's score (51 to 50, the comparison Google leads with) for less than a fifth of the cost per task and in a fifth of the time. At xhigh it ties Carbon at low (54) for 44% of the cost and in about half the time. The model card says Hydrogen is "adapted from Gemini 4 Carbon," so a distilled model matching its parent's cheapest setting is plausible rather than surprising.

Against OpenAI, Hydrogen is a speed play, not a price play. GPT-6.1 Terra at max matches Hydrogen's 51 for about half the cost per task, $0.96 against $1.87, but takes 14.2 minutes to Hydrogen's 6.0. GPT-6.1 Sol at max scores 55 for $1.27, while Hydrogen needs xhigh and $3.16 to reach 54. Sol does it by reading 1.92 billion input tokens, a sixth of what Hydrogen at xhigh reads. Hydrogen is still faster, 13.0 minutes to 22.1.

And xhigh is expensive for what it adds. Going from high to xhigh adds three index points, more than doubles input tokens (5.34B to 11.3B), nearly triples output and more than doubles wall-clock time. Google's own guide says the same in gentler words: reserve xhigh for "hard, multi-step coding and agentic problems where high falls short," because it "can cause overthinking on simpler prompts." The new level exists only on Hydrogen. Carbon and Argon still stop at high and default to medium, while Hydrogen defaults to high, which is worth remembering when comparing "default" results across the family.

What Google's own table shows

Google's announcement says Hydrogen "matches or outperforms competing models in its class." The comparison table in the model card picks Claude Sonnet 5.5 and GPT-6.1 Terra as that class, and Hydrogen does lead on most rows: SWE-sweep (13.5% against 3.1% and 5.8%), Agents' Last Exam v1.1, AutomationBench-AA, Vending-Bench 2, MKA, Chartography and HLE-Diamond. But Sonnet 5.5 beats it on three of the ten rows, and they are the agentic ones developers tend to care about: Terminal-Bench 4.1 (63.9% against 57.0%), OSWorld-2.1 (83.9% against 77.6%) and Terminal-Bench-Science 0.1 (59.9% against 51.4%). Terra also beats it on Terminal-Bench-Science (58.3%). On the Artificial Analysis numbers above, Sonnet 5.5 gets that edge at more than five times Hydrogen's cost per task and five times its time.

A smaller inconsistency: Google's launch post says the new Orchestrator mode in Antigravity, which switches between Carbon and Hydrogen inside one session, reached near-Carbon task success on "Terminal-Bench 4.0" while using 24% less quota and finishing 19% faster. The Antigravity team's post gives the same figures for "Terminal-Bench 4.1." Same numbers, two benchmark versions.

Migration notes

From the What's new in Gemini 4 Hydrogen guide and the model page: the thinking levels are low, medium, high (default) and xhigh, set with thinking_level. Input, thinking and output share one 1,048,576-token window, synchronous responses stop at 65,536 tokens, and longer outputs need background "long decode." As with Carbon, explicit context caching is not supported, only implicit caching. Hydrogen is also now the default model behind the Antigravity agent in Gemini Managed Agents, and it is on the API free tier, which Carbon is not. Google recommends running evaluations "at all four levels" and choosing the lowest one that meets your quality bar. Given the table above, that advice is worth taking literally.

The short version: if you're on Gemini 3.8 Flash for agent work, Hydrogen is the easy call even at today's introductory Flash price, and from January 1 it costs the same per token anyway. If you're choosing across vendors, Hydrogen buys speed. GPT-6.1 Terra and Sol get similar scores for less money, just more slowly.