Google announces Gemini 4 Hydrogen, rolling out to Gemini app

Google unveiled Gemini 4 Hydrogen on November 12, 2026, dropping the Flash moniker to establish a new tier designed for everyday speed and intelligence across consumer and developer platforms. As detailed in 9to5Mac's report, the model succeeds Gemini 3.8 Flash, which launched in September, and introduces native context windows spanning up to one million input tokens alongside one million output tokens. Google positioned the release directly after the rollout of Gemini 4 Carbon, configuring Hydrogen to finish tasks 2.5 times faster under its default high setting while reducing fabricated claims by more than five times.

This technical breakdown examines the operational capabilities, benchmark numbers, application integrations, developer pricing, and system orchestrations delivered with Hydrogen. The release targets everyday mobile users drafting content, enterprise teams automating specialized legal or tax workflows, and developers looking to balance reasoning depth against API token budgets.

01Standalone speed and benchmark performance

Google reported that Gemini 4 Hydrogen more than triples its predecessor's performance on Terminal-Bench 4.1, climbing from 18.2 percent on Gemini 3.8 Flash to 57 percent. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level and reaches 54 at its highest xhigh thinking setting, edging out Argon's score of 50. Outside partners testing the release observed task execution running 2 to 2.8 times faster while overall token consumption dropped between 12 percent and 23 percent.

For multi-step autonomous workloads, Hydrogen ranks fifth overall on arena.ai's live Agent Arena benchmark covering deep research, data analysis, and codebase navigation. Domain-specific evaluations show Hydrogen leading its class on HLE-Diamond at 58.8 percent, ranking first overall on Vals Tax Agent Bench at 52.3 percent, tying for first on Vals Legal Research Bench at 55.8 percent, and leading its class on AutomationBench-AA. It also reached 53.4 percent on the Multimodal Knowledge Agent benchmark and 71.6 percent on Agents' Last Exam v1.1.

02Agent reliability and tool boundaries

Engineering evaluations highlight significant decreases in tool failures and agentic overreach compared to the September 3.8 Flash baseline. Hydrogen cuts agentic overreach by up to 81 percent, under-service by up to 79 percent, and misaligned actions across coding evaluations by twelve times. The system adapts cleanly when an external tool call fails or when a search returns an empty result set, preventing repetitive execution loops.

Safety and guardrail testing shows Hydrogen is up to two times less likely to bypass explicit permission boundaries when confronting blocked or impossible tasks. Google recorded a 96.7 percent honesty rate on internal evaluations, paired with an 80 percent reduction in fabricated assertions. These constraints prevent automated agents from taking shortcuts or misreporting outcomes when operating over complex multi-step workflows.

03Everyday help inside the Gemini app

Inside the consumer Gemini application, Hydrogen takes over the Everyday help tier in Google's revised model lineup. The updated selector organizes Quick answers under 3.5 Flash-Lite, Everyday help under 4 Hydrogen, Detailed research under 3.1 Pro for AI Pro subscribers, Code and analysis under 4 Argon, and Most complex work under 4 Carbon. Free app users receive access to Hydrogen automatically through Auto model selection whenever prompts match the model's profile.

Google tuned Hydrogen to adhere strictly to requested tone, formatting specifications, and negative rules while generating structured data tables and visual breakdowns. The model routes tasks to Connected Apps and specialized skills more frequently, producing ready-to-use study guides, rewrites, and document drafts on initial attempts. Hydrogen also serves as the default engine behind Google Search AI Mode, parsing conversational queries and typos, while powering complex questions in AI Overviews.

04Orchestrator workflows in Google Antigravity

In Google Antigravity, Hydrogen introduces an Extra thinking tier alongside existing Low, Medium, and High settings for paid Google AI Pro and Ultra subscribers. Google also launched Orchestrator, an execution engine that dynamically routes operational steps between Carbon and Hydrogen within a single working session. This architecture preserves Carbon-grade reasoning output while offloading repetitive steps to Hydrogen to prevent excessive token burn.

For collaborative development tasks requiring parallel agents, Carbon serves as the central coordinator while multiple Hydrogen instances run worker and verifier roles simultaneously across isolated workspaces. Teams can execute code generation, build verifications, and testing suites concurrently without inflating deployment expenses. Developers calling the standalone model via API pay $1.50 per million input tokens and $7.50 per million output tokens.

05Structural transition across Google model tiers

Hydrogen's rollout marks the formal phase-out of Flash branding for Google's lightweight flagship tier, standardizing the lineup around elemental designations. The shift aligns consumer mobile chat tools, web search retrieval layers, and enterprise developer tools around a single shared medium-weight model. The architecture targets high-frequency reasoning tasks that previously overwhelmed smaller models but did not justify the operational latency of Carbon.

The combination of one-million-token input and output limits with reduced tool shortcutting establishes a repeatable baseline for multi-agent workflows. With Hydrogen assuming daily duties across Search AI Mode, the Gemini consumer app, and developer agent swarms, Google's stack relies on runtime delegation between specialized tiers rather than routing every request to a single monolithic system.