Google Releases Gemini 4 Hydrogen: 5x Reduction in Hallucinations, Doubling of Speed

Google expands the reach of the Gemini 4 series with a high-performance model aimed at both general users and developers.

Following the launch of its flagship model Carbon earlier this week, Google released Gemini 4 Hydrogen on Thursday, positioning it with the core mission of "bringing cutting-edge intelligence to everyone."

This model focuses on speed and execution efficiency, abandoning the previous "Flash" brand name in favor of Hydrogen.

Google stated that Hydrogen has been globally launched for developers, enterprises, and consumers, covering multiple platforms including Gemini API, AI Studio, Gemini applications, Google Search, and Google Antigravity.

In benchmark tests, Hydrogen surpassed the Argon model in the same series on the Artificial Analysis Intelligence Index with a score of 51 to 50 in default high-thinking mode, reaching 54 points when the highest xhigh thinking mode was activated.

Compared to the previous generation Gemini 3.8 Flash, Hydrogen completed tasks 2.5 times faster under the same settings, with a Terminal-Bench 4.1 score jumping from 18.2% to 57%, and the hallucination generation rate (fabricated claims) reduced by over 5 times.

For developers, the model is priced at $1.5 per million input tokens and $7.5 per million output tokens, with pricing calculated uniformly across the entire 1 million token context window.

Positioning: Efficiency Execution Layer Completes Multi-Model Collaboration Puzzle

Google's division of labor for the Gemini 4 series has become clearer: Carbon and Argon handle deep reasoning and complex tasks, while Hydrogen serves as a fast execution engine suitable for everyday coding, agent workflows, and multi-step data processing.

To systematize this multi-model collaboration model, Google has simultaneously launched the "Orchestrator" feature in Google Antigravity. This feature dynamically schedules Carbon and Hydrogen within the same session, automatically switching based on task complexity without user intervention.

Internal testing data disclosed by Google shows that the Orchestrator achieves a task success rate close to Carbon on Terminal-Bench 4.0, while reducing quota consumption by an average of 24% and increasing speed by 19%.

In parallel agent workflows, Carbon can act as a coordinator, while Hydrogen drives work and validates agents, synchronously building and testing code changes in isolated workspaces—aiming to help teams accelerate complex project delivery while controlling token costs.

Consumer Side: Comprehensive Upgrades for Search and Gemini Applications

On the product level aimed at general users, Hydrogen has become the default model for Google Search AI Mode and is also used for handling the "hardest questions" in AI Overviews.

Google stated that Hydrogen can better understand users' true intentions, even when query statements contain multiple sub-questions, difficult-to-recognize spelling errors, or highly specific needs, allowing the search system to intelligently parse. Additionally, search results will feature more visual presentations and cover a wider range of relevant links.

In the Gemini application, Hydrogen is positioned as a "daily assistant" in the model selector, situated between "quick answers" (3.5 Flash-Lite) and "in-depth research" (3.1 Pro). Free users can utilize Hydrogen through the "[Auto] automatic model selection" mechanism when prompts are adapted.

Google stated that this model can more strictly adhere to the tone, format, and rules set by users, allowing email drafts, study guides, and documents to be directly usable upon first generation.

In Google Antigravity, Hydrogen has added an "Extra" thinking level, which paid AI Pro and Ultra subscription users can access, along with the aforementioned Orchestrator multi-model collaboration capabilities.

Performance: Dual Leap in Speed and Accuracy

Google positions Hydrogen as a "significant independent upgrade" over 3.8 Flash, with core improvements focused on speed, accuracy, and contextual processing capabilities.

In terms of speed, early testing data from partners shows that Hydrogen runs 2 to 2.8 times faster in complex graph calculations, SQL, and data-intensive enterprise workflows, while reducing token usage by 12% to 23%.

Yashodha Bhavnani, VP of enterprise software company Box, stated that Hydrogen significantly improves accuracy while running at about twice the speed of 3.8 Flash, with a 12% reduction in token usage. Yashodha Bhavnani said:

The improvement is most evident in data-intensive workflows that require accurate reading of tables and execution of complex calculations.

In terms of accuracy, Hydrogen supports native input of 1 million tokens and up to 1 million tokens output, capable of handling dense PDF files, spreadsheets, charts, audio, and video content.

In multi-modal knowledge agent (MKA) benchmark tests, Hydrogen scored 53.4%, surpassing other models; in the Agents' Last Exam v1.1 test aimed at end-to-end professional workflows, it scored 71.6%.

In vertical fields, Hydrogen leads peer models with a score of 58.8% in the HLE-Diamond deep domain reasoning test, ranks first with 52.3% in the Vals Tax Agent Bench, and ties for first with 55.8% in the Vals Legal Research Bench.

The team at biotechnology company LatchBio reported that Hydrogen achieved a pass rate of 67% in the toxicology and safety section of its antibody development benchmark test, compared to only 37% for the Gemini model.

Safety: Significant Strengthening of Agent Alignment Capabilities

Google emphasizes the improvements in Hydrogen's agent safety alignment, which is particularly crucial for the autonomous workflow execution scenarios that this model targets.

Compared to 3.8 Flash, Hydrogen reduces "overreach" in agent tasks by up to 81%, "under-service" by up to 79%, and decreases misalignment behaviors and shortcut phenomena in tool invocation and code evaluation by 12 times.

When faced with tasks that cannot be completed or are blocked, the probability of Hydrogen bypassing permission boundaries is reduced by up to 2 times, achieving a 96.7% honesty rate in Google's internal agent honesty assessments.

In terms of reasoning transparency, Hydrogen maintains 99.9% readability of thought chains across over 200,000 audited dialogues, with no records of deceptive reasoning edits.

In the Gray Swan IPI benchmark test against indirect prompt injection attacks, Hydrogen's resistance to attacks has improved by over 2.3 times.

Additionally, Google stated that the updated safety barriers have significantly improved accuracy in identifying real biological safety risks and legitimate medical or scientific issues, with a false rejection rate of less than 0.4% on HealthBench Pro.