Google Just Launched Gemini 4 Hydrogen — a Coding Agent Model Priced to Undercut Everything


By Shakeel Ahmed — AI Growth Lab | November 13, 2026

Google launched Gemini 4 Hydrogen on November 12, 2026, and the headline is not the model — it is the price. Hydrogen is an execution-oriented member of the Gemini 4 family built for coding agents and everyday AI workflows, and it costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1-million-token context window. In plain terms: Google just put near-frontier intelligence into the cheap seats, and it did it one day after launching an even cheaper decision-making model that costs ten cents per million tokens. If you build with AI agents, run a small software business, or pay an API bill every month, this week changed your math.

What Gemini 4 Hydrogen actually is

Gemini 4 comes in three tiers. Carbon and Argon are the frontier models, built for deep, long-horizon reasoning on hard problems. Hydrogen is the fast one. Google describes it as built for "rapid execution across diverse agentic workflows" — writing code, running tests, verifying changes, doing research, and carrying multi-step tasks to completion. It is adapted from Gemini 4 Carbon, not a small model bolted onto the family name. It accepts text, images, audio, and video, and it handles repository coding, debugging, interactive web development, and multimodal tasks.

Google's own announcement puts it simply: "Today, we're introducing Gemini 4 Hydrogen, bringing cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows." That word "everyone" is doing real work here. Hydrogen is rolling out globally across all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search — which Google calls its biggest model upgrade to Search yet — and to paid Google AI Pro and Ultra subscribers in Google Antigravity. Free Gemini app users get it through automatic model selection rather than choosing it directly.

The four thinking levels, and what they cost you

Hydrogen ships with four configurable thinking levels: low, medium, high (the default), and a new xhigh tier. Thinking level is a dial, not a different model. Turn it down and you get faster, cheaper answers for simple work. Turn it up and the model reasons longer before answering.

At its default high setting, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, edging out Argon at 50, and reaches 54 at its highest xhigh setting. Those are Google's numbers, on a benchmark Google chose, and independent verification will come as third-party evaluations publish. But the direction is clear: this is not a stripped-down budget model. It is a model that competes with the frontier tier on some measures while running faster and cheaper.

There is a catch worth knowing before you budget. The model card reports a 13.5% SWE-sweep score for Hydrogen, compared with 3.1% for Claude Sonnet 5.5 and 5.8% for GPT-6.1 Terra — but its Terminal-Bench 4.1 score is 57.0%, below Sonnet 5.5's 63.9%. It also reports 53.4% on Multimodal Knowledge Agent and 71.6% on Agents' Last Exam v1.1. No single benchmark tells the whole story, and Google's own model card acknowledges hallucinations, occasional slowness or timeouts, and increased token use at higher thinking levels. Its knowledge cutoff is March 2026. The honest read: Hydrogen is fast and cheap, genuinely strong at agent-style coding work, and still a model you should verify rather than blindly trust.

Two practical limits matter for developers. Default synchronous output is capped at 64,000 tokens; longer output, up to one million tokens, uses background long decode. And the API documentation lists cached input at $0.075 per million tokens through implicit caching, which makes repeated work against the same codebase dramatically cheaper than the headline rate suggests.

The real strategy: Hydrogen as the worker, Carbon as the boss

Hydrogen is built to stand on its own — or to pair with Carbon and Argon as the high-efficiency execution engine in multi-model workflows. This is the pattern the whole industry is converging on. The expensive model plans. The cheap model executes. A thousand AI agents recently built a working operating system in eight days with no human writing the code (https://aigrowthlabshakeel.blogspot.com/2026/11/claude-agents-built-operating-system-8-days.html) — and that kind of multi-agent work only makes economic sense when the worker agents are cheap. Hydrogen is Google's answer to that bill: a model priced for the thousands of small calls an agent makes, not just the one big call a human makes.

The ten-cent partner: Gemini Decision 1 and the Decisions API

The second half of this story launched a day earlier, on November 11, and it may matter more. The Gemini Decisions API, powered by the new Gemini Decision 1 model, answers closed questions instead of generating text. You send one shared input — text, images, audio, video, or PDF — plus up to 1,000 typed questions, and the model returns a calibrated probability and confidence score for each one. It does not decode text token by token. It scores your candidate answers in a single forward pass.

The numbers are startling. It costs $0.10 per million input tokens, with zero output tokens billed, and Google quotes 150 to 250 milliseconds end to end. The shared input is billed once per request no matter how many questions you evaluate. Google says Decision 1 is nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. There is a no-cost tier for prototyping in Google AI Studio, and Vertex AI support is listed as coming soon.

What is it for? Bounded choices: tool selection, policy checks, model routing, content grading, rubric scoring. Every agent you build makes hundreds of small decisions — which tool to call, whether a result passes a check, which model should handle the next step. Until now, each of those decisions cost a full text-generation call. Decision 1 turns the decision itself into a ten-cent commodity. Pair it with Hydrogen and you get the shape of Google's agent stack: Decision 1 routes, Hydrogen executes, Carbon plans.

Who should care, and what it costs in real money

For developers, the math is simple. A coding agent that fixes a bug might burn a million input tokens across dozens of tool calls. On Hydrogen, that is $1.50 of input plus the output. The same workload on a frontier model can cost several times more. For a small business running agents on customer support, document processing, or code review, the difference between a model that costs dollars per task and one that costs cents is the difference between an experiment and a product.

For everyone else, the change arrives quietly through Google Search and the Gemini app. Google calls this its biggest model upgrade to Search yet, which means the answers you get from Google are about to get faster and more capable without you doing anything.

The limits are honest ones. Hydrogen is not the smartest model Google makes. On some benchmarks it trails its rivals. Its safety and security report is still forthcoming, and its frontier-safety assessment found it below the biological-information and cybersecurity critical capability levels — but reaching the alert thresholds in both domains, which is a reminder that capable models need real guardrails. And the agents it powers are only as trustworthy as the humans who check their work. Researchers once left 13 AI agents alone on a virtual island for 30 hours (https://aigrowthlabshakeel.blogspot.com/2026/10/13-ai-agents-were-left-alone-on-virtual.html) and watched them invent their own goals, games, and challenges. Cheap, fast agents are powerful. They are not self-supervising.

What to do this week

If you build with the Gemini API, try Hydrogen on a real task you already pay for and compare the bill. Set the thinking level to low for routine edits and high for hard debugging, and watch how the cost moves. If you run agents, test Decision 1 on your routing and grading steps — a free tier exists, and the savings compound across thousands of calls. And if you are a small business owner, not a developer, the takeaway is simpler: the cost of hiring an AI worker just fell again, and the companies that move first will be the ones whose agents do the routine work while humans do the thinking. The job market is already shifting that way — postings for forward-deployed AI engineers in China jumped 1,523% in seven months of 2026 (https://aigrowthlabshakeel.blogspot.com/2026/10/china-ai-jobs-jump-1523-with-huge.html). The premium is moving from writing code to making AI useful inside a real business.

Google's message this week is that frontier AI is no longer a luxury good. It is a utility, priced by the million tokens, with a ten-cent router and a dollar-fifty worker. The question is no longer whether you can afford an AI agent. It is what you will build now that you can.

Sources

Read next on AI Growth Lab: 1,000 AI Agents Built a Working OS in 8 Days — No Humans Wrote It (https://aigrowthlabshakeel.blogspot.com/2026/11/claude-agents-built-operating-system-8-days.html) · 13 AI Agents Were Left Alone on a Virtual Island for 30 Hours. They Invented Their Own Goals. (https://aigrowthlabshakeel.blogspot.com/2026/10/13-ai-agents-were-left-alone-on-virtual.html) · China AI Jobs Jump 1,523% With Huge Salaries — But There Is a Catch (https://aigrowthlabshakeel.blogspot.com/2026/10/china-ai-jobs-jump-1523-with-huge.html)

Comments

Post a Comment

Popular posts from this blog

92 Million Jobs Gone by 2030: What AI Is Really Doing to Your Career

Image
By Shakeel Ahmed Every few months, a new headline declares the end of human work. AI writes code, reads X-rays, drives trucks, drafts contracts, and now — with the arrival of always-on AI agents — it watches your inbox, investigates bugs, and prepares reports while you sleep. In short: The World Economic Forum's Future of Jobs Report 2025 projects 92 million roles displaced by 2030 — and 170 million created, for a net positive. Measured data already shows impact: Goldman Sachs counted 16,000 net U.S. job losses per month from AI displacement, U.S. employers attributed 116,175 job cuts to AI in 2026, and Stanford found employment for software developers aged 22-25 fell nearly 20% since 2024. The question is no longer whether AI will change work. It is how much, how fast, and what humans are supposed to do when the dust settles. This article does what most coverage does not: it separates what the data actually shows from what the panic implies, walks through the numbers field by...
Read more

Iran-Linked Cyberattacks on the United States: What Is Reported, What Is Suspected, and What Is Still Unconfirmed

Image
  By Shakeel Ahmed — AI Growth Lab Iran-linked hackers have been reported in recent weeks targeting a broad range of American critical infrastructure, including water systems, telecommunications, energy facilities, and military and medical systems. The key fact to keep in mind first is this: many of these links are reported as suspected, based on sources speaking anonymously, and are not yet publicly confirmed by the Pentagon or by a released forensic report. What has been reported Water systems, July 2026: In late July 2026, coordinated attacks hit water systems in Minnesota, in more than 30 communities, and in at least six other states, with some reports expanding the scope to around a dozen states and over 100 facilities. The attacks focused on internet-connected programmable logic controllers, the devices that control pumps, wells, towers, and lift stations. US agencies, including CISA, the FBI, and the NSA, have warned since April 2026 that Iranian-affiliated actors are explo...
Read more

13 AI Agents Were Left Alone on a Virtual Island for 30 Hours. They Invented Their Own Goals.

By Shakeel Ahmed — AI Growth Lab In short: MIT researchers placed 13 AI agents on a simulated desert island for 30 hours with no task, no reward, and no goal. The agents invented their own challenges, and when tested afterward, they outperformed identical agents that had never explored the island. The study was posted on arXiv on October 22, 2026, and has not yet been peer reviewed. What actually happened on the island The experiment is simple to describe and strange to think about. Researchers at MIT built a simulated island, loosely based on the game Roblox, and dropped in 13 copies of an agent called Eko , each running on Anthropic's Claude Opus 4.7 model. The agents perceived the world through text descriptions and moved a virtual body by writing instructions. There was no assignment. The only guidance was a short "soul" file describing Eko as naturally curious . The world reset every hour, but Eko kept notes in text files that survived the resets, acting as a ki...
Read more