Google Releases Gemini 4 Hydrogen: 2.5x Faster with Million-Token Context Support

Google releases Gemini 4 Hydrogen, targeting agentic scenarios with high-speed inference, 1M context, and adjustable thinking levels.
Google launched Gemini 4 Hydrogen, claiming 2.5x the speed of Gemini 3.8 Flash at $1.50/$7.50 per million input/output tokens. It supports a 1M-token context window and 4 adjustable thinking levels, letting developers trade latency against reasoning depth. Benchmarks: Artificial Analysis score 51, Terminal-Bench 4.1 pass rate 57%. The article notes these reflect specific evaluation sets, not production stability, and the speed claim lacks disclosed test conditions. Google aims to win on cost-performance in high-throughput agentic scenarios, but weights, technical reports, and per-level performance curves remain unpublished. Developers are advised to run small-scale tests on their own high-frequency tasks first.
Google recently launched its next-generation agentic model, Gemini 4 Hydrogen, which focuses on high-speed inference and ultra-long context. The company claims it is 2.5x faster than Gemini 3.8 Flash. The model also continues Google's aggressive pricing strategy, at $1.50 per million input tokens and $7.50 per million output tokens.

Core Specs: An Engineering Focus Built for Agentic Scenarios
Judging from the published specifications, Gemini 4 Hydrogen's positioning is very clear — it serves agentic workflows. The 1M-token context window means it can process ultra-long documents, large codebases, or complex multi-turn conversation histories in a single pass, without frequent truncation or segmentation. For task-oriented agents that need to continuously call tools and maintain state, this directly reduces the complexity of context management.
Another interesting design choice is the "4 thinking levels." The model allows callers to choose different reasoning depths based on task difficulty — using shallower levels for fast responses in latency-sensitive scenarios, and deeper levels for complex reasoning tasks. This adjustable reasoning budget effectively hands the decision of "how to trade off performance against cost" to developers, rather than covering all scenarios with a single tier.
Agentic workflow refers to a mode of operation in which a large model doesn't just passively answer questions, but autonomously plans steps, calls external tools (search, code execution, APIs, etc.), adjusts its strategy based on intermediate results, and continuously maintains task state. Compared with traditional single-turn Q&A, the key characteristics of such workflows are "multi-step" and "stateful": the model needs to repeatedly read and write context within a single task, accumulating intermediate artifacts. Common implementations include ReAct-style think-act loops, function-calling mechanisms with tool use, and long-horizon task agents that need to retain memory across sessions. The demands this places on a model center on three things: the ability to preserve information over long contexts, the parameter accuracy of tool calls, and the stability of each decision step — any failure in one link can cascade into subsequent failures.
Thinking levels (thinking budget / reasoning depth) is a mechanism that lets callers control how much internal reasoning computation the model invests before generating an answer. It is typically implemented by having the model first carry out a reasoning process that isn't directly output, then produce the final answer; the deeper the level, the more reasoning steps the model can unfold, and the higher the accuracy on complex problems tends to be — but latency and token consumption rise in tandem. Shallow levels approximate direct answering and suit simple tasks such as classification, extraction, and format conversion. The engineering significance of this design is that it turns the reasoning budget into an adjustable parameter, allowing developers to dynamically allocate based on task difficulty, SLA requirements, and cost ceilings, rather than being forced to switch between models of different sizes.
Performance Data: The Gap Between Benchmark Scores and Real-World Performance
The two results Google published are an Artificial Analysis score of 51 and a 57% pass rate on Terminal-Bench 4.1. The former is a composite capability evaluation, while the latter is closer to real engineering scenarios — Terminal-Bench examines a model's ability to complete multi-step tasks in a command-line environment, which has strong reference value for agentic applications.
What needs to be viewed soberly is that such benchmark scores reflect performance on specific evaluation sets and cannot be directly equated with stability in production environments. The claim of a 2.5x speed improvement also lacks public disclosure of testing conditions; the actual experience will be affected by factors such as context length, thinking level, and concurrency.
Terminal-Bench is an agentic evaluation benchmark oriented toward command-line environments. Tasks typically require the model to complete multi-step operations in a real shell, such as configuring environments, debugging build errors, and handling files and version control. The evaluation looks not only at final results but also at whether the model can correctly select and combine commands and diagnose and correct itself after errors. Because it simultaneously examines tool use, error recovery, and long-chain task execution, it is widely regarded as closer to real engineering scenarios than static Q&A benchmarks, and therefore has strong reference value for agentic applications. It should be noted that pass rates are affected by task difficulty distribution and differences in runtime environments, so cross-model score comparisons should be made with caution.
Pricing and Competitive Landscape
At $1.50 per million input tokens and $7.50 per million output tokens, the pricing falls in the mid-to-low range among current mainstream models. Combined with 1M context and high-speed inference, Google clearly hopes to establish a cost-performance advantage in the "high-throughput agent" niche. After all, for teams deploying agents at scale, inference costs scale linearly with the number of calls, and differences in unit price quickly amplify into significant operational budget gaps.
However, the information officially disclosed so far is quite limited. Model weights, detailed technical reports, and the specific performance curves corresponding to each thinking level have not yet been made public. Based on a release note alone, it is still difficult to judge its true level on key dimensions such as long-context retrieval accuracy and tool-calling reliability.
Long-context retrieval accuracy refers to a model's ability to locate and use key information within ultra-long inputs, often tested in a "needle in a haystack" style. A nominal 1M-token context window does not mean the model can use all information at that length with equal quality — actual performance often fluctuates with the position of the target information (beginning, middle, end), the density of distractor content, and total length, exhibiting the so-called "lost in the middle" phenomenon. For agentic scenarios, this is especially critical: if key constraints in a long document are missed, subsequent tool-calling decisions will drift. Therefore, when evaluating, it is advisable to use long texts from your own business for position-distribution testing, rather than looking only at the nominal window size.
Observation: The Focus of Competition in Agentic Models Is Shifting
The release of Gemini 4 Hydrogen sends a signal: the center of gravity in model competition is shifting from "whose leaderboard score is higher" to "who is better suited to running agents." Context length, reasoning levels, per-token price, and terminal task completion rate — the combination of these metrics says more about whether a model can handle real production work than a composite evaluation score alone.
For developers, such models are worth watching but there is no need to rush into migration. A more pragmatic approach is to first run small-scale comparative tests using your own business's high-frequency tasks, focusing on information recall quality under long contexts and the failure rate of multi-step tasks, before deciding whether to add it to your tech stack.
One-line Summary
Google releases Gemini 4 Hydrogen, targeting agentic scenarios with high-speed inference, 1M context, and adjustable thinking levels.
Paragraph Summary
Google launched its next-generation agentic model Gemini 4 Hydrogen, which it claims is 2.5x faster than Gemini 3.8 Flash and priced at $1.50 per million input tokens and $7.50 per million output tokens. The model supports a 1M-token context window and 4 adjustable thinking levels, allowing developers to trade off latency against reasoning depth based on task difficulty. On benchmarks, it scores 51 on Artificial Analysis and has a 57% pass rate on Terminal-Bench 4.1. The article argues that such metrics reflect performance on specific evaluation sets and cannot be directly equated with production stability, and that the speed figures lack disclosure of testing conditions. Google clearly hopes to establish a cost-performance advantage in high-throughput agentic scenarios, but weights, technical reports, and performance curves for each level have not been made public. The article suggests developers first run small-scale comparative tests with their own high-frequency tasks, focusing on long-context recall quality and multi-step task failure rates, before deciding whether to migrate.
Related articles

OMLX 0.7.1 Integrates Splash: Major Speedup for MLX Model Inference on Mac
OMLX 0.7.1 adds Splash as an external engine, letting standard MLX models (Qwen 3.6 35B) match Splash custom setups at 92.8 tokens/s. Setup, D-Flash2 requirements, and M1/M2 compatibility explained.

Running Uncensored AI Locally on Your Phone: LFM3 Hands-On Test and Guide
Learn how to run Liquid AI's LFM3 1.7B uncensored model locally on iPhone and Android—no GPU or network needed. Includes On Device LLM and Layla steps, speed tests, and privacy benefits.

Claude Launches Collaborative Sheets and Whiteboards, Taking On Microsoft and Google Directly
Anthropic rolls out beta Claude Sheets and Whiteboards to paid users, enabling one-prompt spreadsheets with formulas and turning verbal plans into collaborative boards—taking on Microsoft and Google directly.