Google’s Gemini 4 Hydrogen Costs Twice as Much Per Token as Gemini 3.8 Flash but Finishes Agent Tasks for About a Quarter of the Price, Benchmark Data Shows

Google’s newest AI model looks more expensive on paper and cheaper in practice. Gemini 4 Hydrogen, released on November 12, lists at exactly double the per-token price of the Gemini 3.8 Flash model it replaces. Yet independent benchmarking firm Artificial Analysis found that Hydrogen completed its test workloads for about a quarter of the cost per task, because it finishes faster and burns fewer tokens getting there. For any business budgeting for AI agents, that gap between sticker price and real cost is the story.

What Google Launched

In its launch announcement, Google DeepMind positioned Hydrogen as the fast, efficient workhorse of the Gemini 4 family, sitting below its frontier models Gemini 4 Carbon and Gemini 4 Argon. Google says Hydrogen is built for “rapid execution” in everyday coding and agent workflows, while Carbon and Argon handle the deepest, long-horizon reasoning.

The headline claims from Google:

• Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default “high” thinking setting, edging past Argon’s 50, and reaches 54 at a new “xhigh” setting.
• Compared with Gemini 3.8 Flash, it finishes tasks 2.5 times faster, more than triples Terminal-Bench 4.1 scores (57% versus 18.2%) and cuts fabricated claims by more than five times.
• Early partners running data-heavy enterprise workflows saw it run 2 to 2.8 times faster while using 12% to 23% fewer tokens than 3.8 Flash.
• It ships with a 1 million-token context window and can generate up to 1 million output tokens.

Hydrogen is rolling out across Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. Free Gemini app users get it through automatic model selection.

The Price Tag Versus the Real Bill

According to Google’s Gemini API pricing page, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, with thinking tokens billed as output. The rate is flat across the full 1 million-token window, with no long-context surcharge. Cached input drops to $0.075 per million tokens through automatic caching, a 95% discount.

Gemini 3.8 Flash currently lists at $0.75 and $3.75, half of Hydrogen’s rate. But Artificial Analysis’ head-to-head comparison shows why the per-token number misleads. On its test suite, Hydrogen averaged $1.87 per task against $7.77 for 3.8 Flash. Running the firm’s full Intelligence Index cost $1,768 on Hydrogen versus $5,764 on Flash. Hydrogen averaged 7.9 minutes per task compared with 14.7 minutes, and produced output at 179 tokens per second versus 112.

The reason is simple: a smarter model takes fewer wrong turns. Artificial Analysis measured about 85,000 output tokens per task for Hydrogen versus roughly 98,900 for Flash. A model that needs fewer retries, fewer reasoning loops and fewer failed tool calls can end up far cheaper, even at twice the unit price.

There is also a calendar factor. A pricing comparison from DataStudios reports that Gemini 3.8 Flash’s standard rate rises to $1.50 and $7.50 on January 1, 2027, the same list price as Hydrogen. If that holds, Flash’s price advantage on standard requests lasts only about seven more weeks.

Where Flash Still Wins

Hydrogen is not an automatic swap for every workload. The DataStudios comparison notes that Google has not yet published batch, flex or priority pricing for Hydrogen, and that the new model supports only implicit caching, not the explicit context caches some applications rely on. Overnight batch jobs, apps built around explicit caches and traffic that needs a guaranteed priority tier may be better left on Flash for now. Artificial Analysis also measured a slightly longer mean end-to-end time for Hydrogen on single requests (68.4 seconds versus 54.3), a reminder that “faster per task” and “faster per reply” are not always the same thing.

Google’s Bet on Multi-Model Workflows

Alongside Hydrogen, Google introduced Orchestrator in its Google Antigravity development environment. Instead of using one model from start to finish, Orchestrator switches between models like Carbon and Hydrogen in the same session. Google says that on Terminal-Bench 4.0, Orchestrator came close to Carbon’s success rate while using 24% less quota and finishing 19% faster. In parallel setups, Carbon can act as the coordinator while Hydrogen powers the worker and verifier agents.

That design mirrors how many companies already run human teams: senior staff set direction, and faster, cheaper resources do most of the execution. Expect rivals to push the same “router plus workhorse” pattern, because it lets vendors sell frontier quality without frontier prices on every token.

Why It Matters for Business

For finance and operations leaders, the lesson is to budget AI by outcome, not by token. Comparing rate cards will push teams toward the cheapest-looking model, which may cost several times more once retries, longer runs and human cleanup are counted. The better metric is cost per completed task: per resolved support ticket, per reconciled invoice, per qualified lead.

Reliability matters as much as price. Google says Hydrogen cuts overreach in agent tasks by up to 81% and is up to twice as unlikely to bypass permission boundaries when a task is blocked. Those are vendor-reported figures that buyers should test on their own data, but they point to where competition is heading: models that can be trusted to act inside business systems without constant supervision.

Practical steps for teams on older Gemini models: run a small pilot on real workloads, track cost and time per finished task rather than per token, keep batch and explicit-cache jobs where they are until Google publishes Hydrogen’s terms, and revisit the plan before January 1.

The Bottom Line

Gemini 4 Hydrogen shows that the cheapest AI model on a price sheet is not always the cheapest model to run. Benchmark data puts its per-task cost at roughly a quarter of its predecessor’s despite a doubled token price. As AI agents take on more daily work, companies that measure what each finished job really costs will get far more from their AI budgets than those watching the rate card.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous