Gemini 4 Hydrogen replaces Flash at $1.50/$7.50: faster agents, triple the Terminal-Bench score, and a higher bill per token

Google released Gemini 4 Hydrogen on November 12, the third model in the Gemini 4 family and the successor to Gemini 3.8 Flash as Google’s default for everyday coding and agent work. It is live in the Gemini API, Google AI Studio and Android Studio as gemini-4-hydrogen, is the new default model in AI Mode in Search, and reaches free Gemini app users through automatic model selection when Google judges a prompt suits it.

Two things make it more than a routine “faster, cheaper” refresh. On Google’s own numbers it roughly triples Flash’s score on an agentic terminal benchmark and edges past Gemini 4 Argon on an independent intelligence index. And it is not cheaper: per token, it costs twice what Flash costs today. Two days after Carbon reached the API, as we covered in our look at Carbon’s launch and pricing, the Gemini 4 lineup now has a clear shape: Carbon for hard, long-horizon reasoning, Argon for most engineering work, and Hydrogen for high-volume execution.

The price is Flash’s future price, starting now

Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens (reasoning tokens included), flat across its full context window, with cached input at $0.075. VentureBeat’s breakdown supplies the comparison that matters: Gemini 3.8 Flash is on an introductory rate of $0.75/$3.75 until December 31, then moves to $1.50/$7.50 on January 1, 2027. So Hydrogen doubles today’s Flash price but matches what Flash will cost in seven weeks. Batch, Flex and the invite-only Contributor tier run at half the standard rate; Contributor requires letting Google use request data for training.

Google’s deprecation schedule lists no shutdown date for Flash, so teams on cheap Flash traffic are not forced to move yet. They should still plan for January, when the reason to stay on Flash for price alone disappears.

Against the low end of the market, Hydrogen is not a budget model. VentureBeat puts it at 15 times the lowest standard per-token rates of Claude Haiku 5.5, GPT-6 Luna and Meta’s Muse Ember. The argument for it is cost per finished task, not cost per token.

What the benchmarks say, and what they cost

The headline figures, mostly vendor-reported:

Measure	Gemini 4 Hydrogen	Comparison
Artificial Analysis Intelligence Index (high)	51	Argon 50, Flash 31
Same index at xhigh / medium / low	54 / 41 / 27
Terminal-Bench 4.1 (Google’s run)	57%	Flash 18.2%, Claude Sonnet 5.5 63.9%
HLE-Diamond	58.8%
Vals Tax Agent Bench	52.3% (ranked first)
Agents’ Last Exam v1.1	71.6%

Google also says Hydrogen finishes tasks about 2.5 times faster than Flash at default settings and fabricates claims more than five times less often. Box CEO Aaron Levie reported agent accuracy on Box’s Complex Work Eval rising from 74% on Flash to 79% on Hydrogen, in under half the time and with about 12% fewer tokens.

The more useful numbers are Artificial Analysis’s estimated cost per benchmark task: $0.29 at low thinking, $0.80 at medium, $1.87 at high and $3.16 at xhigh, against $7.77 for Flash at high. That is the real case for the higher token price: Flash spent far more tokens to reach a much lower score. Note one inefficiency, though. On Artificial Analysis’s Terminal-Bench 4.0 run, Hydrogen scored 54% at both high and xhigh, so the jump to xhigh added about $1.29 per task for no gain on that benchmark. Default to high, measure, and only raise the level where your own evaluations show it pays.

Two caveats: Hydrogen still trails Claude Sonnet 5.5 on Terminal-Bench 4.1, and cheaper models remain competitive on cost. Xiaomi’s MiMo-V2.6-Flash runs about $0.13 per task at a 24-point lower index score.

Migration details developers will hit

The model card and API notes contain several breaking or behavior-changing details:

• Default thinking is now high. Flash defaulted to medium, so a drop-in swap will spend more reasoning tokens unless you set the level. The four levels are low, medium, high and the new xhigh.
• thinking_budget and MINIMAL are rejected. Code that tuned Flash with a token budget or the minimal setting needs to switch to named levels.
• Context and output. The 1M-token window is shared by input, output, reasoning and tool calls. Synchronous output defaults to 64K tokens; supported APIs can go to 1M through background long decode.
• Caching. Automatic caching on Google Cloud starts at 4,096 tokens (Flash: 6,144). Explicit caching is not available at launch, which matters for anyone whose cost model depends on it.
• Knowledge cutoff is March 2026, the same as Carbon. The model card says Hydrogen is “adapted from Gemini 4 Carbon” and defers to Carbon’s documentation for architecture, data and hardware.

Google is also pushing a two-model pattern. In Antigravity, a new Orchestrator switches between Carbon and Hydrogen within one session: Carbon coordinates while Hydrogen runs worker and verifier agents. Google says that mix came close to Carbon’s Terminal-Bench 4.0 success rate while finishing 19% faster and using 24% less quota. That planner-plus-cheap-workers design is how most production agent stacks already control spend, and it is the strongest hint about where Google expects Hydrogen’s volume to come from.

The safety paperwork: no new frontier risk, more over-refusal

Hydrogen ships with a model card, but its full Safety and Security Report is listed as “coming soon,” the same gap Carbon had at announcement. The card does contain the Frontier Safety Framework result. Google ran its Material Capability Change test against Argon, the last model to get a full critical capability assessment, and found no material increase, so a full evaluation suite was not required. That is consistent with the index numbers: Hydrogen beats Argon by one point, not by a frontier leap.

Because Google could not rule out that Argon and Carbon reached the critical capability levels for biological information risk and cyber uplift, it ran targeted tests on Hydrogen in both domains. Hydrogen did not reach either critical level but did cross the alert threshold for both, which is Google’s early-warning tier. The other risk findings are carried forward from Argon.

The automated content safety table is a mixed result Google publishes plainly. Against Flash, Hydrogen improves image-plus-text safety by 7.00 points, text safety by 0.80 and multilingual by 1.10. But unjustified refusals on benign or borderline prompts rose 5.57 points, “particularly for multimodal prompts,” and refusal tone slipped 0.65 points. Google says manual review found the safety losses were mostly false positives, but the refusal regression is the one product teams will feel: an app that sends images to Flash may see more declined requests after switching. The launch post separately claims agentic overreach down by up to 81%, permission-boundary bypasses down by up to half, and 2.3 times more resilience on Gray Swan’s indirect prompt injection benchmark.

What to watch

• January 1. When Flash’s introductory pricing ends, Hydrogen at the same token price becomes the obvious default for most Flash workloads, and Flash’s remaining role will be latency-sensitive work that does not need reasoning.
• The safety report. Its detail on the cyber and biology alert thresholds will show how close a mass-market, free-tier model now sits to Google’s critical levels.
• Refusal behavior in the wild. The 5.57-point unjustified-refusal increase is an automated measure; developer reports over the next few weeks will show whether it is visible in real multimodal apps.
• Independent agent benchmarks. Most of Hydrogen’s agent claims come from Google and launch partners. Third-party runs on coding agents and tool-use suites will test the 2.5x speed and cost-per-task story.