Google Launches Gemini 4 Hydrogen at $1.50 and $7.50 per Million Tokens, Double Gemini 3.8 Flash's Introductory Price
Google says its new fast model for coding and agents finishes tasks 2.5 times faster than Gemini 3.8 Flash and scores 51 on the Artificial Analysis index, but it costs more per token than rivals.

Key facts
- Google launched Gemini 4 Hydrogen on 12 November 2026, rolling it out across the Gemini API, AI Studio, Gemini for business, the Gemini app and Google Search.[1][3]
- Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its 1-million-token context window.[1][2]
- Gemini 3.8 Flash costs $0.75 and $3.75 per million input and output tokens through 31 December 2026, then $1.50 and $7.50 from 1 January 2027, according to VentureBeat.[2]
- Gemini 4 Hydrogen scores 51 on the Artificial Analysis Intelligence Index at high reasoning, against 31 for Gemini 3.8 Flash.[1][2]
- Google reports Hydrogen scores 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash, while Claude Sonnet 5.5 scores 63.9% on the same table.[2][3]
Google released Gemini 4 Hydrogen on 12 November 2026, a fast model for coding and agent work priced at $1.50 per million input tokens and $7.50 per million output tokens, twice the current introductory rate of the Gemini 3.8 Flash model it succeeds. Google says Hydrogen finishes tasks about 2.5 times as fast as Flash at its default setting, and independent benchmarks put it well ahead on intelligence but not on cost per token.
The launch, described in Google's announcement, moves the company's mid-tier workhorse closer to its frontier models. It also tests whether developers will pay more for a model that finishes work sooner.
What did Google release?
Hydrogen is positioned below Gemini 4 Carbon and Gemini 4 Argon, which Google reserves for deep, long-horizon reasoning. Unite.AI reported that Hydrogen is adapted from Carbon and aimed at fast execution such as coding, testing and tool-using agents. It accepts text, images, audio and video and shares a 1,048,576-token window across input, reasoning and output.
Rollout began globally on 12 November across the Gemini API, Google AI Studio, Gemini for business, the Gemini app and Google Search. Free users of the Gemini app receive it through automatic model selection when Google judges it suitable. Google AI Pro and Ultra subscribers get it in the Antigravity coding tool, and VentureBeat reported that Google Cloud lists the model as generally available.
Hydrogen has four reasoning levels: low, medium, high, which is the default, and a new xhigh tier. Google's blog post says the levels scale in a smooth, predictable progression.
How much does it cost?
The price is the story's sharpest edge. According to VentureBeat, Hydrogen costs $1.50 per million uncached input tokens, $0.075 per million cached input tokens and $7.50 per million output tokens, including reasoning tokens. Gemini 3.8 Flash is on an introductory rate of $0.75 and $3.75 through 31 December 2026, rising to $1.50 and $7.50 on 1 January 2027. Hydrogen therefore matches Flash's scheduled post-introductory rate and is double the current one.
VentureBeat also reported that batch, flex and the invite-only contributor tier are each half the standard rate, and that Hydrogen is about 15 times the lowest standard rates of Claude Haiku 5.5, GPT-6 Luna and Meta's Muse Ember. Google says the price is flat across the full context window.
Price per token is not the same as price per task. Using Artificial Analysis estimates relayed by VentureBeat, Hydrogen at high reasoning costs about $1.87 per task, against about $7.77 for Gemini 3.8 Flash at high reasoning. Smaller rivals are cheaper still: Haiku 5.5 at maximum reasoning is estimated at $0.41 per task and Meta's Muse Ember at $0.14.
How does it perform?
On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at high reasoning and 54 at xhigh, according to Google and VentureBeat. Gemini 3.8 Flash at high scores 31. Google's post says the Gemini 4 Argon model scores 50.
On Terminal-Bench 4.1, a test of command-line agent work, Google reports 57% for Hydrogen against 18.2% for Flash. Both VentureBeat and Unite.AI note that Claude Sonnet 5.5 scores higher on the same table, at 63.9%. Most figures originate with Google or its partners. Box chief executive Aaron Levie posted a chart, relayed by VentureBeat, showing accuracy on the Box Complex Work Eval rising from 74% for Flash to 79% for Hydrogen, with about 12% fewer tokens used per typical task.
One result complicates the case for the highest setting. VentureBeat reported that on Terminal-Bench 4.0, Hydrogen scored 54% at both high and xhigh, while the estimated cost per task rose from $1.87 to $3.16. Google itself recommends xhigh only for harder multistep problems, noting that it can overthink simpler prompts, according to Unite.AI.
Google also reports Antigravity's new Orchestrator, which lets a session move between Carbon and Hydrogen, reached close to Carbon's task success while using about 24% less quota and finishing about 19% faster. Unite.AI said Carbon plans in a read-only phase and Hydrogen executes, with control returning to Carbon if execution runs into trouble.
What should developers watch for?
VentureBeat listed several migration changes. Hydrogen defaults to high reasoning, adds xhigh, and rejects the `thinking_budget` parameter and the `MINIMAL` setting. On Google Cloud, explicit caching is unavailable at launch, and automatic caching starts at 4,096 tokens, down from 6,144 on Flash. Because the roughly one-million-token window is shared by input, reasoning and output, higher reasoning settings also consume room that a long prompt might otherwise use.
Unite.AI, citing Google's model card, reported a March 2026 knowledge cutoff and noted that unjustified refusals rose 5.57 percentage points compared with Gemini 3.8 Flash in the card's automated comparisons. The same model card says Hydrogen stayed below the critical capability levels for biological and cybersecurity risk but reached alert thresholds in both areas, and Google says fuller safety documentation is forthcoming. Google's own safety claims, including up to 81% fewer cases of agentic overreach than Flash, come from internal evaluations.
What we don't know yet
Independent verification is thin. Most benchmark numbers come from Google, its partners or Artificial Analysis figures reported second-hand, and real-world cost depends on how much reasoning a workload needs. It is also unclear how many developers will move from cheaper rival models, or whether Google will adjust Hydrogen's pricing when Flash's introductory rate ends on 31 December.
Why it matters
Hydrogen shows a shift in how AI labs sell their fast tiers. The older bargain was a cheap model that was good enough; the newer pitch is a model that costs more per token but finishes sooner and wastes fewer tokens. If Google's figures hold up, the cost per completed task may fall even as the price list rises, which would put pressure on rivals that compete mainly on rate cards. If they do not, developers running high volumes of simple calls have cheaper alternatives, as VentureBeat's comparison table shows.

