Gemini 4 Hydrogen Released: A Daily Execution Engine Combining Speed and Intelligence

Gemini 4 Hydrogen targets daily coding and agentic workflows with high speed and low cost, complementing the deep-reasoning flagship.
Google DeepMind launches Gemini 4 Hydrogen, a high-efficiency execution model for everyone, complementing Carbon and Argon's long-horizon reasoning. It uses flat-rate pricing ($1.50/$7.50 per million input/output tokens) and scores 51 on the Artificial Analysis Intelligence Index. Antigravity's Orchestrator dynamically switches models, nearing Carbon's success rate on Terminal-Bench 4.0 while cutting quota use by 24% and boosting speed 19%. Migrating from 3.8 Flash brings 2.5x faster task completion, Terminal-Bench 4.1 rising from 18.2% to 57%, and over 5x fewer fabricated claims. Hydrogen offers four thinking levels, million-token input/output, and significantly reduces over-action and prompt injection risks.
Google DeepMind has officially launched Gemini 4 Hydrogen, positioned as a high-efficiency execution model that "brings frontier intelligence to everyone." Unlike Gemini 4 Carbon and Argon, which focus on deep long-horizon reasoning, Hydrogen is purpose-built for everyday coding and agentic workflows, emphasizing a balance between speed and efficiency.



Product Positioning: Not the Flagship, but More Practical
According to Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at Google DeepMind, Hydrogen's goal is to extend Gemini 4's capabilities across all tiers—from Google AI Studio and the Gemini API to the Gemini app, Google Search, and paid subscribers. This means it is not merely an experimental model for developers, but a daily engine embedded directly across Google's full product line.
The pricing strategy reflects this positioning as well: across the entire 1 million token context window, input is priced at $1.50 per million tokens and output at $7.50 per million tokens, at a flat rate that does not increase with context length.
On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at the default high thinking level, slightly above Argon's 50; it can reach 54 at the highest xhigh thinking setting. Although the lead is modest, combined with its speed advantage, the difference in experience during actual workflows will be more noticeable.
The Artificial Analysis Intelligence Index is a composite benchmark maintained by the independent evaluation organization Artificial Analysis. It aggregates multiple standardized tests (covering dimensions such as reasoning, knowledge, and coding) into a single score for easy cross-model comparison. An important feature of this index is that it allows for differences in "thinking level"—the same model will receive different scores under different reasoning budgets. Therefore, when the article mentions that Hydrogen scores 51 at the default high thinking level and can reach 54 at the xhigh setting, this range reflects performance fluctuation of the same model when reasoning depth is adjustable, rather than a fixed gap between different models. When interpreting such scores, readers should pay attention to the thinking settings and cost assumptions behind them, and avoid simplistic horizontal comparisons.
Multi-Model Collaboration: Orchestrator Automatic Scheduling
Google has introduced Orchestrator in Antigravity, which can dynamically switch between models such as Carbon and Hydrogen within the same session. On Terminal-Bench 4.0, Orchestrator achieves task success rates close to Carbon while reducing quota consumption by an average of 24% and increasing completion speed by 19%. For scenarios requiring parallel agents, Carbon can serve as the coordinator while Hydrogen acts as worker and verifier, simultaneously building and testing changes in isolated workspaces.
This division of labor addresses a real-world problem: teams need deep reasoning models when designing systems or troubleshooting complex bugs, but when writing code, running tests, and verifying changes, a fast and disciplined execution model is more suitable. Hydrogen can independently handle daily work and also serve as a high-efficiency execution engine in multi-model workflows.
Terminal-Bench is a benchmark for agents executing tasks in real terminal environments. It examines whether a model can complete multi-step tasks such as file operations, code modifications, and environment configuration through the command line, emphasizing end-to-end task success rates rather than single-turn question-answering quality. Quota consumption refers to the quota units consumed by a task under subscription or API billing systems, usually tied to token usage and reasoning depth. The reason Orchestrator can reduce quota consumption by 24% while achieving success rates close to Carbon is that it allocates deep reasoning to the few steps that truly require it, while delegating a large number of repetitive execution and verification steps to efficient models like Hydrogen, thereby compressing the share of expensive reasoning overall. Understanding these two concepts helps grasp the trade-off logic between cost and effectiveness in multi-model scheduling.
Performance Leap: Significant Gains for Teams Migrating from 3.8 Flash
For teams migrating from Gemini 3.8 Flash, Hydrogen delivers noticeable upgrades. At the default high setting, task completion speed increases by 2.5x, the Terminal-Bench 4.1 score rises from 18.2% to 57%, fabricated claims decrease by more than 5x, and it adapts more cleanly when tool calls fail or searches return empty results.
For iterative repository coding and debugging, Hydrogen excels at focused modifications and avoids rewriting unrelated files. It can natively execute file reads and search calls in parallel batches rather than inspecting files one by one. In early testing, partners saw Hydrogen run 2x to 2.8x faster on complex graph, SQL, and data-intensive workflows, while token usage decreased by 12% to 23%.
In open agentic execution, Hydrogen ranks 5th on the Agent Arena live benchmark; leads peers on HLE-Diamond at 58.8%; ranks 1st on the Vals Tax Agent Bench at 52.3%; ties for first place on the Vals Legal Research Bench (55.8%); and leads peers on AutomationBench-AA.
In multimodal knowledge work, Hydrogen supports native 1 million token input and up to 1 million token output, and can synthesize dense PDFs, spreadsheets, charts, audio, and recorded video into cited reports and structured deliverables. It scores 53.4% on the Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1.
Fabricated claims refer to content that a model generates that appears plausible but is actually untrue when lacking supporting evidence—a common failure mode for agents in tool calls, retrieval, and report writing. Reducing the number of fabricated claims means the model is more inclined to answer based on actual tool return results rather than filling in gaps from thin air. HLE-Diamond is a high-difficulty subset of Humanity's Last Exam, used to examine the upper limits of a model's knowledge and reasoning on expert-level questions; Agent Arena is a leaderboard that evaluates agent task performance through live comparisons, and its rankings change with participating models and task distribution. These benchmarks each have different emphases: some measure depth of knowledge, others measure agentic execution capability, and a single score is insufficient to judge a model's overall performance in real workflows.
Four Thinking Levels: Allocate Reasoning Depth on Demand
Hydrogen offers four thinking levels—low, medium, high (default), and the newly added xhigh—forming a smooth and predictable progression curve among quality, latency, and cost. Users only wait and pay for the reasoning depth the task requires. Across thinking levels, Hydrogen matches or exceeds competing models of the same class, and typically completes faster with fewer output tokens.
Consumer Experience: Upgrades in the Gemini App and Search
In the Gemini app, Hydrogen is better at following user instructions and calling the right tools, reducing back-and-forth communication. It closely adheres to tone, format, and specific rules, making email drafts, study guides, rewrites, and documents usable on the first pass. When comparing products or synthesizing dense materials, it can generate clear tables and visual breakdowns without losing key details.
In Search, Hydrogen brings three improvements: better understanding of real intent, even when users have multiple query components, typos, or highly specific requests; more adaptive, visual, and scannable responses; and connection to a broader range of relevant links through upgraded query fan-out technology.
Safety and Alignment: Built for Large-Scale Daily Execution
Hydrogen inherits the Gemini 4 safety stack, reducing over-action in agentic workflows by up to 81%, reducing under-action by up to 79%, and reducing misaligned behavior and task shortcuts by up to 12x in tool-use and coding evaluations. When facing impossible or blocked tasks, the likelihood of bypassing permission boundaries is reduced by up to 2x, and it achieves a 96.7% honesty rate in internal agent honesty evaluations.
In terms of reasoning transparency, Hydrogen maintains 99.9% chain-of-thought readability across more than 200,000 audit turns, with zero deceptive reasoning edits. Against indirect prompt injection attacks, resilience on the Gray Swan IPI benchmark improves by more than 2.3x. In daily interactions, updated safeguards more precisely distinguish genuine biosecurity risks from legitimate medical or scientific questions, keeping the false rejection rate on HealthBench Pro below 0.4%.
Indirect Prompt Injection (IPI) is an attack method targeting agentic systems: attackers hide malicious instructions in external content that the model will read, such as web pages, documents, or tool return results, inducing the agent to deviate from its original goal, leak information, or perform unauthorized operations during task execution. Unlike malicious prompts directly entered by users, IPI is more covert because the legitimate user is not malicious. The Gray Swan IPI benchmark is used to measure a model's resilience against such attacks. Over-action and under-action are a pair of concepts in agent alignment evaluation; the former refers to the model performing operations beyond authorization or necessity, while the latter refers to the model failing to execute or executing insufficiently when it should. Reducing these two types of deviations is key to making agents both reliable and non-transgressive in daily workflows.
Availability: Immediate Full Rollout
From launch day, Hydrogen is rolling out globally: developers can use it through the Gemini API in AI Studio and Android Studio paid tiers; paid Google AI Pro and Ultra subscribers can try it in Google Antigravity. Enterprise users can use it in Gemini for business, Google Antigravity, and eligible Google Sheets business subscriptions. All users can access it in the Gemini app model dropdown and Google Search, while Gemini features in Google Sheets start with AI Pro and Ultra subscribers.
Hydrogen's release sends a clear signal: competition in frontier intelligence is not only about benchmark scores, but also about whether it can integrate into the daily workflows of hundreds of millions of users at acceptable cost and speed. For most developers and ordinary users, a model that is smart enough and responsive is often more valuable than a flagship that is occasionally stunning but slow to respond.
One-line Summary
Gemini 4 Hydrogen enters daily coding and agentic workflows with a high-speed, low-cost positioning, complementing the deep-reasoning flagship.
Paragraph Summary
Google DeepMind has launched Gemini 4 Hydrogen, positioned as a high-efficiency execution model for everyone, forming a division of labor with Carbon and Argon, which focus on long-horizon reasoning. It uses flat-rate pricing ($1.50 per million input tokens, $7.50 per million output tokens) and scores 51 on the Artificial Analysis Intelligence Index at the default high thinking level. The newly added Orchestrator in Antigravity can dynamically switch models within the same session, reducing quota consumption by 24% and increasing speed by 19% while approaching Carbon's success rate on Terminal-Bench 4.0. After migrating from 3.8 Flash, task completion speed increases by 2.5x, the Terminal-Bench 4.1 score rises from 18.2% to 57%, and fabricated claims decrease by more than 5x. Hydrogen offers four thinking levels, supports million-token input and output, and significantly reduces over-action and prompt injection risks in safety alignment.
Related articles

AI Agent Orchestration Framework Tier List: Who's Overrated, Who's Legendary
A blunt tier-list evaluation of AI Agent orchestration frameworks: why CrewAI and LangGraph are legendary, OpenAI Agents SDK and Agno are only good, LlamaIndex and Cloudflare Agents are overrated, and why TypeScript developers should choose Mastra.

Giving AI Agents Memory: OpenAI SDK + Mem0 in Practice
Use the OpenAI Agent SDK with Mem0 to add persistent memory to AI agents: from installing dependencies and configuring credentials to creating a memory client and isolating users by ID, build a memory-enabled conversational agent in three steps.

Svelte: The Compile-Time Framework That Makes Web Development Simple Again
Svelte is a compile-time frontend framework with 89k+ GitHub stars, no virtual DOM, and tiny bundles. This article covers its core differences, the SvelteKit ecosystem, use cases, and drawbacks.