Google Deploys Gemini 4 Hydrogen to Search Globally, Adding Third Variant to Gemini 4 Series

Google DeepMind released Gemini 4 Hydrogen on November 12 [4] and integrated it into Google Search worldwide [5][6], marking the third model in the Gemini 4 family after Argon and Carbon. The company positions Hydrogen as a faster, lower-cost daily driver for agentic workflows while its two siblings target deeper reasoning tasks.

A Three-Model Family in About a Month

Gemini 4 Argon launched on October 13 across the Gemini API, AI Studio, the Gemini app and AI Mode in Search [1]. Gemini 4 Carbon followed with a restricted preview for cyber defenders in late October and rolled out broadly on November 10 via the Gemini API, Google AI Studio, Google Antigravity, and the Gemini app for AI Ultra subscribers [3]. Hydrogen now arrives as the speed-optimized sibling, available immediately for all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity [5].

Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, authored the announcement [5]. Robby Stein, VP of Product for Google Search, published a companion post detailing the Search integration [6].

Search Integration: Global Rollout for All Users

Google Search is receiving Hydrogen globally for all users starting today. The company says the model improves intent understanding for complex, multi-part queries, expands "query fan-out" to surface a wider range of links, and delivers more visual responses including structured tables and custom tools [6]. Free Gemini app users will access Hydrogen via automatic model selection when the system deems it appropriate [5].

Pricing and Developer Access

Google's official announcement describes Hydrogen's pricing as flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens [5]. That undercuts Argon's introductory $2/$10 rates (which double to $4/$20 after November 30) [1]. The model writes up to 64,000 tokens in a single reply and offers four thinking levels—low, medium, high (default), and xhigh [4]. The announcement also mentions an output context window of up to 1 million tokens [5]; the 64,000-token single-reply limit comes from third-party documentation [4].

Third-party API provider aihubmix.com lists a "Cache Read" price of $0.075 per million tokens for Hydrogen, which equals a 95% discount on the standard input rate [2]. Google's announcement does not mention a cached-input discount for Hydrogen, creating a discrepancy between the official statement and the provider's published pricing.

Google also introduced Orchestrator in Google Antigravity, a system that dynamically switches between Carbon and Hydrogen within a single session. On Terminal-Bench 4.0, Google says Orchestrator reached near-Carbon success rates while using 24% less quota and completing 19% faster [5].

Benchmark Claims: Speed vs. Depth

Google reports Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level (versus Argon's 50), rising to 54 at xhigh [5]. It claims 76.1% on AutomationBench and 58.8% on HLE-Diamond, beating Claude Sonnet 5.5 and GPT-6.1 Terra on those tests [4][5]. Output speed is listed at 243 tokens/second, though time-to-first-token sits at 29.1 seconds on a log scale—significantly slower than Claude Sonnet 5.5 (1.0s) and GPT-6.1 Terra (4.3s) [4].

The company says Hydrogen finishes tasks 2.5x faster than Gemini 3.8 Flash at its high default setting and more than triples Terminal-Bench 4.1 scores (57% vs. 18.2%) [5]. Partner testimonials from Cartwheel, LatchBio, Box, Pixelcut, and Astrocade are quoted in the announcement but have not been independently verified [5].

Safety Posture

The model card shows Hydrogen did not reach Critical Capability Levels (CCL) for CBRN biological, cybersecurity, or harmful manipulation uplift, though alert thresholds were reached for biological and cybersecurity [8]. For the Loss of Control – Stealth and Situational Awareness TCL, Google says it is applying mitigations as if the threshold had been reached "as a conservative assumption" [8]. For the CBRN biological and cybersecurity CCL alert thresholds, Google says it continues to deploy mitigations [8]. The company claims 99.9% chain-of-thought legibility with zero deceptive reasoning edits across 200,000+ audited turns, and a 96.7% honesty rate on internal agent evaluations [5].

Compressed Cadence, Open Questions

Three Gemini 4 models in about a month reflects an accelerated release rhythm. Argon shipped broadly on day one; Carbon rolled out in stages; Hydrogen launches globally across consumer and developer surfaces simultaneously. Source [3] describes Google as rapidly building out a full periodic table of specialized models and notes leadership has previously referenced Hydrogen alongside whispers of Krypton and Xenon [3], though the actual gaps—28 days from Argon to Carbon's broad release, then two days to Hydrogen—are uneven.

What remains unclear: whether Carbon will reach general availability before Argon's price doubling on November 30 [1], and how the three-model lineup will be presented to non-technical users in the Gemini app. Google's blog says free users get Hydrogen via auto-selection "when it's the appropriate model for the prompt" [5], but does not define that threshold.

The Search integration is the widest deployment yet for a Gemini 4 model. Whether Hydrogen's faster-but-shallower profile improves Search quality at scale—or simply changes the error distribution—will take weeks of real-world usage to assess.