Google is broadening the reach of its Gemini 4 lineup with a high-efficiency model aimed at both everyday users and developers.
Following the debut of its flagship Carbon model earlier this week, Google launched Gemini 4 Hydrogen on Thursday, positioning it around the core idea of "bringing frontier intelligence to everyone."
The model focuses on speed and execution efficiency, abandoning the long-used "Flash" branding in favor of the name Hydrogen.
Google said Hydrogen is now available simultaneously worldwide for developers, enterprises and consumers, spanning multiple platforms including the Gemini API, AI Studio, the Gemini app, Google Search and Google Antigravity.
On benchmarks, Hydrogen outperforms the same series' Argon model in the default high-thinking mode on the Artificial Analysis Intelligence Index (51 points versus 50 points), and reaches 54 points when the highest xhigh thinking mode is enabled.
Compared with the previous generation Gemini 3.8 Flash, Hydrogen completes tasks 2.5 times faster under equivalent settings, lifts its Terminal-Bench 4.1 score from 18.2% to 57%, and reduces the rate of fabricated claims by more than five times.
For developers, the model is priced at US$1.5 per million input tokens and US$7.5 per million output tokens, with flat pricing across the entire 1 million-token context window.
Positioning: An efficiency execution layer that completes the multi-model collaboration puzzle
Google's division of labor within the Gemini 4 series has become increasingly clear: Carbon and Argon handle deep reasoning and complex tasks, while Hydrogen serves as a fast execution engine suited to everyday coding, agent workflows and multi-step data processing.
To systematize this multi-model collaboration model, Google simultaneously introduced an "Orchestrator" feature within Google Antigravity. The feature dynamically dispatches Carbon and Hydrogen within the same session, automatically switching based on task complexity without requiring manual intervention from users.
Internal test data disclosed by Google shows that Orchestrator approaches Carbon's task success rate on Terminal-Bench 4.0 while cutting quota consumption by an average of 24% and boosting speed by 19%.
In parallel agent workflows, Carbon can act as coordinator while Hydrogen drives work and verification agents, simultaneously building and testing code changes in isolated workspaces—aimed at helping teams accelerate delivery of complex projects while keeping token costs under control.
Consumer side: Search and the Gemini app receive a full upgrade
On the consumer product side, Hydrogen has become the default model for Google Search's AI Mode and is also used to handle the "hardest questions" in AI Overviews.
Google said Hydrogen can better understand users' true intent, so that even when a query contains multiple sub-questions, hard-to-decipher spelling errors or highly specific requirements, the search system can parse it intelligently. In addition, search results will be presented more visually and cover a broader range of relevant links.
Within the Gemini app, Hydrogen is positioned in the model selector as the "everyday assistant," sitting between "quick answers" (3.5 Flash-Lite) and "deep research" (3.1 Pro). Free users can access Hydrogen through the "[Auto] automatic model selection" mechanism when their prompts fit.
Google said the model follows user-specified tone, format and rules more strictly, allowing email drafts, study guides and documents to be usable directly upon first generation.
In Google Antigravity, Hydrogen adds an "Extra" thinking tier, available to paying AI Pro and Ultra subscribers, along with access to the aforementioned Orchestrator multi-model collaboration capability.
Performance: A dual leap in speed and accuracy
Google positions Hydrogen as a "major standalone upgrade" over 3.8 Flash, with core improvements concentrated in three dimensions: speed, accuracy and context handling.
On speed, early partner test data shows Hydrogen runs 2 to 2.8 times faster in complex graph computation, SQL and data-intensive enterprise workflows, while using 12% to 23% fewer tokens.
Yashodha Bhavnani, a VP at enterprise software company Box, said Hydrogen delivers significantly better accuracy while running at roughly twice the speed of 3.8 Flash and using 12% fewer tokens. Yashodha Bhavnani said:
The improvements are most pronounced in data-intensive workflows that require accurately reading tables and performing complex calculations.
On accuracy, Hydrogen supports native 1 million-token input and up to 1 million-token output, and can handle dense PDF files, spreadsheets, charts, audio and video content.
On the multimodal knowledge agent (MKA) benchmark, Hydrogen scored 53.4%, beating other models; on the Agents' Last Exam v1.1 test for end-to-end professional workflows, it scored 71.6%.
In vertical domains, Hydrogen leads peer models with 58.8% on the HLE-Diamond deep-domain reasoning test, ranks first with 52.3% on the Vals Tax Agent Bench, and ties for first with 55.8% on the Vals Legal Research Bench.
The team at biotech company LatchBio said Hydrogen achieved a 67% pass rate on the toxicology and safety portion of its antibody development benchmark, compared with just 37% for non-Gemini models.
Safety: Agent alignment capabilities substantially strengthened
Google placed heavy emphasis on Hydrogen's improvements in agent safety alignment, which is especially critical for the autonomous workflow execution scenarios the model targets.
Compared with 3.8 Flash, Hydrogen reduces "overreach" in agent tasks by up to 81% and "under-service" by up to 79%, while misaligned behavior and task shortcuts in tool calling and code evaluation fall 12-fold.
When facing tasks that cannot be completed or are blocked, Hydrogen is up to 2 times less likely to bypass permission boundaries, achieving a 96.7% honesty rate in Google's internal agent honesty evaluation.
On reasoning transparency, Hydrogen maintained 99.9% chain-of-thought readability across more than 200,000 audited conversations, with no records of deceptive reasoning edits.
On the Gray Swan IPI benchmark targeting indirect prompt injection attacks, Hydrogen's resistance improved by more than 2.3 times.
In addition, Google said its updated safety guardrails show markedly improved precision in distinguishing genuine biosecurity risks from legitimate medical or scientific questions, with a false-refusal rate below 0.4% on HealthBench Pro.
Comments