Clarity for AI buildersSponsored by AI Routing
availability

Google releases Gemini 4 Hydrogen, edging Argon on intelligence index with lower-cost positioning

Google releases Gemini 4 Hydrogen, edging Argon on intelligence index with lower-cost positioning
Google has released Gemini 4 Hydrogen, which edges Argon on one intelligence index, The Decoder reports. The general-purpose model is intended for fast, everyday work rather than the deep, long-horizon reasoning targeted by Gemini 4 Carbon and Gemini 4 Argon. According to The Decoder’s report, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default “high” thinking level, compared with Argon’s 50. Hydrogen reaches 54 at its highest setting, “xhigh.” That result gives Hydrogen a narrow advantage over Argon on this index, not a demonstrated lead across every workload. The report’s benchmark table shows Hydrogen ahead on some measures, including Agents’ Last Exam v1.1, but behind Argon on Chartography. Its positioning as an everyday model therefore matters as much as the headline intelligence score: the evidence supports a workload-specific comparison rather than a universal replacement claim. Four thinking levels, different operating costs Hydrogen offers low, medium, high and xhigh thinking levels, while Argon offers medium and high. The reported API prices are $1.50 per million input tokens and $7.50 per million output tokens, with cache hits listed at $0.075. The model supports a one-million-token input context window and up to one million output tokens. Artificial Analysis’s evaluation, as reported by The Decoder, puts average task costs at $0.29 for low, $0.80 for medium, $1.87 for high and $3.16 for xhigh. The corresponding intelligence-index scores are 27, 41, 51 and 54. The xhigh setting uses nearly twice as many output tokens as high. Those figures make the thinking setting a consequential deployment choice. Moving from high to xhigh adds three index points, while the reported average task cost rises from $1.87 to $3.16. Conversely, the cheaper settings give up substantial measured capability. A single model name or token-price listing does not capture that trade-off; the setting and workload are necessary parts of the comparison. Reported gains over Flash, with terminal-work limits Google describes Hydrogen as a substantial upgrade over Gemini 3.8 Flash. According to the report, Google says Hydrogen completes tasks 2.5 times faster at its default thinking level, raises Terminal-Bench 4.1 performance from 18.2% to 57%, and reduces hallucinations by a factor of five. Google also claims better recovery from failed tool calls and empty search results. The independent evaluation provides a more qualified picture. The Decoder reports that Artificial Analysis measured an 8% hallucination rate on AA-Omniscience and found Hydrogen stronger than other Gemini models, including Carbon, on presentation quality in AA-Briefcase. But Hydrogen scored 54% on Terminal-Bench 4.0 and 50% on Terminal-Bench-Science 0.1, trailing Claude Sonnet 5.5 at its maximum thinking setting on both tests. The terminal results should not be treated as interchangeable with Google’s Flash comparison. The reported 57% result uses Terminal-Bench 4.1, whereas the independent 54% result uses Terminal-Bench 4.0. Likewise, the 8% hallucination figure applies to AA-Omniscience, not to every task Hydrogen might perform. These distinctions limit how broadly the improvements can be interpreted. Rollout includes a mixed-model coding workflow The Decoder reports a global rollout through the Gemini API in AI Studio and Android Studio, Gemini for Business, the Gemini app and Google Search. Paying Pro and Ultra subscribers can access Hydrogen through Google Antigravity. Google also says the model follows tone, formatting and instructions more closely during longer Gemini app conversations and improves handling of complex or misspelled Search queries. In Antigravity, a new orchestrator switches between Carbon and Hydrogen within a session. Google says this approach nearly matches Carbon’s success rate on Terminal-Bench 4.0 while using 24% fewer resources and reducing processing time by 19%. That orchestration result concerns a combined workflow, not Hydrogen operating alone. It points to a different practical use of the new model: assigning work across models rather than selecting one model for an entire session. The supplied report does not define the resource measure or provide enough testing detail to establish whether those savings would carry over to other coding workloads.