Google replaces Gemini 3.8 Flash with Gemini 4 Hydrogen
Google has replaced Gemini 3.8 Flash with Gemini 4 Hydrogen across its products. Google claims a 2.5x speed gain and a more-than-5x cut in fabricated claims over Flash 3.8.

- Google replaced Gemini 3.8 Flash with Gemini 4 Hydrogen across its products, with immediate access from rollout.
- In high mode, Hydrogen runs 2.5 times faster than Flash 3.8, per Google's claims.
- Google reports Hydrogen reduces fabricated claims by more than 5x compared with Flash 3.8.
- Hydrogen more than tripled Flash 3.8's score on Terminal-Bench for code generation.
- Artificial Analysis scored Hydrogen 54 on its Intelligence Index in xhigh mode, versus 50 for Argon.
Google has replaced Gemini 3.8 Flash with Gemini 4 Hydrogen across its products, and access began immediately at rollout. The swap ended the staged, paywalled releases that shaped Google's two previous model launches.
Hydrogen replaces the Flash tier. Flash models carried the faster, cheaper workloads across Google's ecosystem. The Pro/Flash naming scheme is gone. Google now labels models with elements: Argon, Carbon, and now Hydrogen.
The rollout history matters here. Argon, announced in September 2026, launched without public access. Carbon, announced in October 2026, followed the same pattern. Both rollouts began with cybersecurity researchers before general availability. Even then, Google limited access to subscribers of its most expensive AI subscription tier. Hydrogen breaks that pattern. The limits are relaxed. Access opened immediately.
What performance does Google claim?
Google claims three concrete gains over Flash 3.8. Hydrogen runs in four thinking levels: low, medium, high, and a new xhigh. In high mode, Hydrogen runs 2.5 times faster than Flash 3.8. Google's own analysis reports a reduction in "fabricated claims" of more than 5x.
Code generation improved the most. Hydrogen more than tripled 3.8 Flash's score on Terminal-Bench, a coding benchmark. That figure comes from Google, not from an independent evaluation. Google has not published the underlying test conditions.
These are vendor claims. Treat them accordingly. Independent benchmark data exists for the intelligence index, but the speed and hallucination figures rest on Google's internal analysis.
Where does Hydrogen land against Argon?
Hydrogen sometimes matches Argon, Google's frontier-tier model released earlier. Artificial Analysis, an independent model-testing firm, scored Hydrogen at 54 on its Intelligence Index in xhigh mode. Argon scored 50. A budget-tier replacement outscoring a frontier model by four points is unusual.
Caveat: that single index does not settle the comparison. The source reports no benchmark breakdowns beyond the composite score. Argon may retain advantages in domains the index does not test. The scores also reflect xhigh mode, Hydrogen's most expensive thinking level, not its default setting.
Why does the naming change matter?
The elemental scheme — Hydrogen, Carbon, Argon — tells users less than the old Pro/Flash labels did. Those names signaled tier and cost. Element names signal nothing about performance class. A reader cannot infer from the name whether a model targets frontier workloads or cheap bulk inference.
The practical signal sits in the rollout policy, not the name. Hydrogen's immediate, unrestricted availability marks it as the volume model. Argon and Carbon remain gated. For anyone building on Google's stack, Hydrogen is now the default assumption.
What should you watch next?
Monitor three things. First, whether independent labs replicate Google's 5x hallucination-reduction claim on real-world queries — vendor analyses routinely flatter. Second, whether the relaxed access policy holds, or Google tightens rate limits once usage scales. Third, whether the Artificial Analysis gap holds across other benchmarks, since a 54-versus-50 result on one index is thin evidence that the Flash tier closed the frontier gap.
via arstechnica.com (Source)
More from Elena Vasquez
Show full bio
Correspondent covering people, launches and day-to-day practice in Search rankings and algorithm updates, generative engine optimization at SERP Journal.
221 articles