Google Claims Gemini 4 Hydrogen Runs 2.5x Faster Than 3.8 Flash

Google DeepMind introduced Gemini 4 Hydrogen on November 12, 2026 as a fast execution model for everyday coding and agentic workflows. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, positions Gemini 4 Carbon and Argon as the deep-reasoning frontier, while Hydrogen is the high-efficiency execution tier—either as a standalone upgrade for teams on Gemini 3.8 Flash or paired with Carbon and Argon in multi-model workflows. The announcement is less a new reasoning ceiling than a repricing and latency play aimed at workloads that fail on cost, turnaround, or tool-calling reliability.

Hydrogen has four thinking levels: low, medium, high as default, and xhigh. Google DeepMind reports that at default high it scores 51 on the Artificial Analysis Intelligence Index, ahead of Argon's 50, and reaches 54 at xhigh, while often finishing tasks faster than models in its class. Pricing is flat across the entire 1 million-token context window: $1.50 per million input tokens and $7.50 per million output tokens.

For multi-model routing, Google introduces Orchestrator in Google Antigravity. Google reports that Orchestrator reached near-Carbon task success on Terminal-Bench 4.0 while consuming 24% less quota and completing 19% faster on average. In parallel-agent workflows, Carbon can coordinate while Hydrogen powers worker and verifier agents in isolated workspaces, separating deep architectural judgment from disciplined execution.

Against Gemini 3.8 Flash, Google reports Hydrogen at its high default setting finishes tasks 2.5 times faster, raises Terminal-Bench 4.1 from 18.2% to 57%, cuts fabricated claims by more than 5x, and recovers cleanly when a tool call fails or a search returns empty. Google also says it outperforms similar-tier models on SWE-sweep and natively batches file-read and search calls in parallel. In early partner testing across graph, SQL, and data-heavy enterprise workloads, Google reports Hydrogen ran 2x to 2.8x faster while cutting token usage 12% to 23% versus 3.8 Flash.

On agentic benchmarks, Google reports Hydrogen ranks #5 on Agent Arena, leads its class on HLE-Diamond at 58.8%, ranks #1 overall on Vals Tax Agent Bench at 52.3%, ties for #1 on Vals Legal Research Bench at 55.8%, and leads its class on AutomationBench-AA. For multimodal knowledge work, Google reports 53.4% on the Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1, with native 1 million-token input and up to 1 million output tokens. All benchmark figures are vendor-reported.

In Search, Google says Hydrogen better parses multi-part, typo-heavy, and highly specific queries, and widens query fan-out so users see links they might previously not have discovered.

Hydrogen rolls out November 12 across all Google AI Studio tiers, the Gemini API, Gemini for business, Google Search for all users, the Gemini app, and paid Google AI Pro and Ultra subscribers in Google Antigravity. Free Gemini app users get Hydrogen through auto model selection, not direct selection, according to the announcement footnote. Gemini in Google Sheets starts with eligible business subscriptions and AI Pro/Ultra subscribers.

Pricing places Hydrogen between Google's budget Flash tier and the previous Pro Preview tier. The comparison below uses AI Mastery's token-calculator data as of 2026-10-22 for the comparison rows; the Hydrogen figures are Google's and are not independently confirmed.

AI Mastery analysis

Hydrogen's practical consequence is for teams that have resisted multi-model orchestration because routing overhead and per-model pricing were unpredictable. Flat $1.50/$7.50 pricing across the full 1M-token context simplifies cost modeling, but the larger source of savings in Google's Orchestrator numbers—24% less quota and 19% faster on Terminal-Bench 4.0—depends on routing accuracy, a system-level property rather than a model property. If the orchestrator sends low-level tasks to Carbon or deep-debugging tasks to Hydrogen, the cost and latency upside disappears. That aligns with our prior finding that production AI fails on architecture, not model intelligence.

The benchmark deltas are best read as another vendor-selected evaluation frame. The Terminal-Bench 4.1 comparison, the reported more-than-5x drop in fabricated claims, and the safety figures—up to 81% less overreach, up to 79% less under-service, a 12x cut in misaligned actions and task shortcutting, 96.7% honesty, 99.9% chain-of-thought legibility, and over 2.3x stronger resistance on the Gray Swan IPI benchmark—are Google DeepMind's own evaluations. Teams should reproduce them on their own traces before relying on the margins for unsupervised agent deployment. Free Gemini app users also cannot force Hydrogen for a specific query style; they depend on Google's routing heuristics.

The broader signal is that model releases are now segmented by execution economics and tool-call reliability as much as by reasoning ceiling. Hydrogen reinforces multi-model pipelines where routing, quota, and task-type tradeoffs are first-class engineering decisions. Technical buyers should treat the vendor benchmark deltas as hypotheses to validate against their own workload traces, not as independent evidence.

Sources

• Primary: Google DeepMind: Introducing Gemini 4 Hydrogen
• Competitor prices: AI Mastery token calculator, figures as of 2026-10-22