Google Releases Gemini 4 Hydrogen: 2.5x Faster Than 3.8 Flash, $1.50 per Million Tokens
Google has released Gemini 4 Hydrogen, a fast model for everyday coding and agent work. It completes tasks 2.5x faster than Gemini 3.8 Flash and scores 57% on Terminal-Bench 4.1 (terminal tasks), up from 18.2%. A million input tokens cost $1.50, and a million output tokens cost $7.50. That makes it cheaper to hand routine work like edits and tests to Hydrogen and keep the complex tasks with Carbon and Argon.
AI-processed from Habr AI; edited by Hamidun News
Large models think well, but they are slow and expensive, and that is overkill for routine work like editing code, running tests, and checking changes. Google's answer is Gemini 4 Hydrogen, built for fast agent tasks (programs that work through a task step by step on their own) and for daily coding work. It runs 2.
5x faster than its predecessor, Gemini 3.8 Flash. On Terminal-Bench 4.
1, a benchmark for command-line tasks, its score rose from 18.2% to 57%. A token is the unit of text you pay for: a million input tokens cost $1.
50, and a million output tokens cost $7.50. The senior models in the family, Gemini 4 Carbon and Gemini 4 Argon, remain the choice for deep, long reasoning over the most complex tasks.
All figures below come from Google and its partners. The article includes no independent verification.
What changed compared with Gemini 3.8 Flash
For teams running Gemini 3.8 Flash, Hydrogen looks like a significant upgrade. At the standard high reasoning level, it completes tasks 2.5x faster. The rate of fabricated claims, meaning cases where the model invents facts, fell by more than five times. It also handles tool-call errors and empty search results better.
The price is the same across the full 1-million-token context window, so a long document does not make a request more expensive per token. The model accepts up to a million input tokens and returns up to a million output tokens. On the Artificial Analysis Intelligence Index, it scores 51 at the standard high reasoning level, compared with 50 for Argon, and 54 at the maximum xhigh level.
Partners report similar numbers. Box got roughly double the speed and 12% fewer tokens. Pixelcut ran 276 trials: the model completed all design tasks, ran 1.8x faster, and wrote 30% cleaner code. STAN cut latency by 28% and token usage by 23%. In early tests on complex graphs and SQL, partners saw 2–2.8x speedups and 12–23% token savings.
Orchestrator: the big model decides, the small one executes
The most practical part of the release is Orchestrator in Google Antigravity. It switches between models within a single session, for example between Carbon and Hydrogen. It works like a team where an architect makes the hard decisions and an executor writes code and runs tests. On Terminal-Bench 4.0, Orchestrator came close to Carbon's task completion rate. It used 24% less quota on average and ran 19% faster.
The pattern also works with parallel agents. Carbon acts as coordinator, while Hydrogen manages executor agents and verifier agents. They create and test changes at the same time in isolated workspaces. According to Google, this speeds up delivery of complex projects without increasing token spending.
There are four reasoning levels: low, medium, high (the default), and the new xhigh. The higher the level, the deeper the model thinks and the longer the user waits. You pay only for the depth the task needs.
Agents, documents, and where the model is available
In Agent Arena on arena.ai, Hydrogen ranks fifth. In its class, it leads on HLE-Diamond (58.8%) and AutomationBench-AA. It also ranks first on the Vals Tax Agent Bench (52.3%) and shares first place on the Vals Legal Research Bench (55.8%). On the new Multimodal Knowledge Agent test, it scored 53.4%, and on Agents' Last Exam v1.1, 71.6%. It can turn PDFs, spreadsheets, diagrams, audio, and video into reports with source citations.
In the Antibody Development test, LatchBio reported that Hydrogen succeeded on 67% of attempts in the toxicology and safety section, compared with 37% for the best model outside the Gemini family, while spending about 13 minutes per task.
The model is available in Google AI Studio, the Gemini API, Gemini for business, the Gemini app, and Google Search. Paid AI Pro and Ultra subscribers get it in Google Antigravity.
The practical takeaway is simple. If your agent spends many tokens on repetitive work like edits, tests, and reading repositories, it is worth testing Hydrogen as the primary executor. Leave complex architectural decisions and difficult debugging to the senior model. Google compiled these benchmarks itself, so before switching, run the model on your own tasks and compare completion rate, time, and token cost.
- Nov 21, 2026AGENTS.md помог Codex соблюдать правила проекта: 0 нарушений из 13 против 5 из 13
- Nov 21, 2026Claude and Codex Took a Blog's CTR from 0.4% to 0.9% in 3 Months and Delivered 1 Million Impressions
- Nov 21, 2026One in Four Programmers Already Works With AI Agents: T1 Survey of 190 Developers
- Nov 21, 2026Meta's Muse Agent Booked an Airbnb Unprompted and Quoted Bank Balances Users Didn't Recognize
Key Questions Answered
What changed compared with Gemini 3.8 Flash?
For teams running Gemini 3.8 Flash, Hydrogen looks like a significant upgrade. At the standard high reasoning level, it completes tasks 2.5x faster. The rate of fabricated claims, meaning cases where the model invents facts, fell by more than five times. It also handles tool-call errors and empty search results better. The price is the same across the full 1-million-token context window, so a long document does not make a request more expensive per token.
What are the key details on Orchestrator: the big model decides, the small one executes?
The most practical part of the release is Orchestrator in Google Antigravity. It switches between models within a single session, for example between Carbon and Hydrogen. It works like a team where an architect makes the hard decisions and an executor writes code and runs tests. On Terminal-Bench 4.0, Orchestrator came close to Carbon's task completion rate. It used 24% less quota on average and ran 19% faster. The pattern also works with parallel agents.
What are the key details on Agents, documents, and where the model is available?
In Agent Arena on arena.ai, Hydrogen ranks fifth. In its class, it leads on HLE-Diamond (58.8%) and AutomationBench-AA. It also ranks first on the Vals Tax Agent Bench (52.3%) and shares first place on the Vals Legal Research Bench (55.8%). On the new Multimodal Knowledge Agent test, it scored 53.4%, and on Agents' Last Exam v1.1, 71.6%. It can turn PDFs, spreadsheets, diagrams, audio, and video into reports with source citations.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.