Google's Gemini 4 Hydrogen Undercuts Argon on Price and Edges It on One Index
Google released Gemini 4 Hydrogen on 12 November 2026 at a flat $1.50 per million input tokens, with a 51 on Artificial Analysis at its default thinking level.

On this page
Key takeaways
- Google announced Gemini 4 Hydrogen on 12 November 2026, a cheaper, faster model for everyday coding and agent workflows.
- It costs a flat $1.50 per million input tokens and $7.50 per million output tokens across a 1 million-token context window.
- Artificial Analysis scores Hydrogen at 51 on its Intelligence Index at the default high setting, one point above Gemini 4 Argon and below Gemini 4 Carbon.
- Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash and more than triples its Terminal-Bench 4.1 score, from 18.2% to 57%.
- Free Gemini app users reach Hydrogen through automatic model selection, not by picking it directly from the model menu.
Google released Gemini 4 Hydrogen on 12 November 2026, the third model in its Gemini 4 family and the one built for ordinary work rather than the hardest problems. The pitch is narrow and concrete: a cheap, fast model for coding, document handling and agent loops, priced at a flat $1.50 per million input tokens and $7.50 per million output tokens. Google added the model to its Gemini line, then spent most of the announcement explaining why a model that is not the smartest is the one most customers should use.
Related reading: Anthropic's Claude Duet Works in the Background, and Asks Before It Acts · An Anthropic Mathematician Says Claude Found a Riemann Hypothesis Proof.
That is a change in emphasis for Google. Gemini 4 Argon arrived in September as the frontier model, and Gemini 4 Carbon reached general sale on 10 November as the expensive reasoning option. Hydrogen is the volume product. It sits below Carbon on the hardest benchmarks, above Argon on one widely watched index at its default setting, and below both on price.
Where Hydrogen sits in the Gemini 4 line
Google describes Carbon and Argon as models for deep, long-horizon reasoning. Hydrogen is for "rapid execution across diverse agentic workflows", the company said in its launch post. In practice that means writing and testing code, reading files, calling tools, checking its own output and moving to the next step.
The model has four thinking levels: low, medium, high and a new xhigh tier. High is the default. Each level trades latency and token spend for reasoning depth, and Google says the progression is meant to be predictable so a developer can pick the cheapest level that clears a task.
The context window is 1 million tokens of input and up to 1 million tokens of output. Pricing does not change with context length, which matters for the kind of work Hydrogen is meant to do. A long repository read costs the same per token as a short prompt.
Google is also using Hydrogen to bring Gemini 4 to every tier of AI Studio and to all users of the Gemini app and Search. Carbon and Argon were paid or gated products at launch. Hydrogen is the first Gemini 4 model Google says free app users can reach, though only through automatic selection.
The price is flat, and the cost per task is the real number
The list price is easy to state. Input costs $1.50 per million tokens, output $7.50, and there is no introductory rate that doubles later. Carbon costs $2 and $10 until 30 November, then $4 and $20. Argon carries the same introductory structure.
Per token, Hydrogen is 25% cheaper than Carbon's launch price and 62.5% cheaper than Carbon's standard price. That comparison flatters Hydrogen less than the per-task number, because a model that thinks longer can erase a lower token rate.
Artificial Analysis puts the estimated cost of one Intelligence Index task at $1.87 for Hydrogen at high. Wccftech, reading Google's comparison table, put xhigh at about 32% of Claude Sonnet 5.5 at max for near-equal intelligence, and about 2.5 times the cost of GPT-6.1 Sol at max. Those are benchmark-suite estimates, not invoices, but they describe the gap Google is selling.
The lower price comes with a harder limit. Hydrogen is not the model for a long debugging session that needs a plan across a dozen files. Google says teams should pair it with Carbon or Argon when the work needs architectural judgment, and use Hydrogen as the execution engine once the plan exists.
That division of labour is the core of the pitch. It is also where the cost argument can break. If a cheap model needs three attempts where an expensive one needs one, the cheap model loses. Google's answer is Orchestrator, a feature in Antigravity that switches between models inside a session.
The benchmark claims, and who ran them
Google reports that Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high setting and 54 at xhigh, against 50 for Argon. Carbon, by contrast, scored 59 when it launched. The one-point edge over Argon is real on that index, and small enough that it should not be read as a generational jump.
Artificial Analysis lists 51 for Hydrogen at high. Vals places it 6th on its Vals Index at 68.49% accuracy, with a measured cost of $9.84 per test and a latency of about 32 minutes per full run. Vals ranks it first of 85 on its Tax Agent Bench at 52.3%, just ahead of Carbon at 51.78% and Claude Fable 5.1 at 50.75%, and second of 40 on its Finance Index.
The vendor numbers are larger. Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at the high setting and raises Terminal-Bench 4.1 from 18.2% to 57%. It reports a 5x reduction in fabricated claims, 58.8% on HLE-Diamond, a tie for first on the Vals Legal Research Bench at 55.8%, 53.4% on the new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1.
Not every result points the same way. Vals places Hydrogen 11th of 64 on Terminal-Bench 4.0 at 54.04% and 13th of 128 on its Vibe Code Bench at 87.78%, which is respectable but not a clean sweep. On the Vals Index it trails Claude Fable 5.2 and sits in the middle of the top group.
The pattern is consistent with the product. Hydrogen is strong for its price, not the strongest model available. Google's own comparison table shows it leading models in its class, a phrase that does work the benchmarks cannot. The class is defined by price and speed, and Google chose the class.
Independent reruns matter here because Google's launch charts are the vendor's own. Artificial Analysis and Vals have both published results, which is more than many launches get, but neither has reproduced every claim in the announcement. The 2.5x speed figure and the 5x fabrication reduction come from Google's internal testing.
Orchestrator is the more interesting product
The multi-model pitch is where Hydrogen becomes more than a cheaper endpoint. In Google Antigravity, Orchestrator can start a session on Carbon and hand execution to Hydrogen, or run Carbon as a coordinator while Hydrogen powers worker and verifier agents in parallel.
Google says that on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success while using 24% less quota and finishing 19% faster on average. Those numbers are the company's, and quota is not the same as dollars. Still, the design addresses a real problem: most agent work is not one hard reasoning step, it is many small steps around one.
The same logic drives Google's parallel-agent example. Carbon plans, Hydrogen builds and tests in isolated workspaces, and the coordinator checks the results. If the cheap workers fail, the expensive model retries. If they succeed, the customer pays the cheap rate for most of the run.
The risk is operational complexity. A team that already struggles to manage one model now has to decide which model handles which step, watch token spend across both, and debug failures that may come from the handoff rather than the model. Google's answer is to automate the choice, which moves the decision into a product the customer cannot fully inspect.
Safety, distribution and the free-tier catch
Google frames Hydrogen as safer for agent work than 3.8 Flash. It says the model reduces overreach by up to 81% and under-service by up to 79%, cuts misaligned actions and task shortcutting by 12x across tool-use and coding evaluations, and is up to twice as likely to respect permission boundaries when a task is blocked. It reports a 94% honesty rate on an internal evaluation and 99.9% chain-of-thought legibility across more than 200,000 audited turns.
The prompt-injection claim is more specific. Google says Hydrogen is more than 2.3x more resilient than 3.8 Flash on the Gray Swan indirect prompt injection benchmark. That matters for a model that reads web pages and files, and it is still a benchmark result from a vendor with a commercial interest in the number.
Distribution is the other half of the launch. Developers get Hydrogen in AI Studio, the Gemini API and Android Studio. Enterprises get it in Gemini for business, Antigravity and Gemini in Sheets for eligible subscriptions. Consumers get it in the Gemini app and in Google Search, which Google says uses Hydrogen to improve intent handling and query fan-out.
The free tier is narrower than the headline suggests. Free Gemini app users do not select Hydrogen from the model menu. They get it through automatic model selection when Google's system decides it is the right model for the prompt. Paid AI Pro and Ultra subscribers can choose it directly in Antigravity and Sheets. The model is available to everyone, but not equally.
What the launch does not establish
What Google has not published
Nearly every performance figure in the Hydrogen announcement comes from Google or a partner. Artificial Analysis and Vals have published independent index results, but they do not cover every claim, including the 2.5x speed gain and the 5x reduction in fabricated claims. Google published no adoption, revenue or retention data. The free tier is automatic model selection rather than direct access. And the model card lists the usual limits: hallucinations, variable latency and higher token use at higher thinking levels.
The price comparison also needs care. A flat $1.50 rate is not a flat cost per job. Hydrogen's default thinking level spends tokens on reasoning, and xhigh spends more. A customer who leaves the model on xhigh for every request will pay more than the headline rate suggests, and a customer who drops to low may get weaker answers.
The competitive picture is crowded. OpenAI's GPT-6.1 Sol and Anthropic's Claude Sonnet 5.5 sit in the same price band, and both have their own cheap tiers. Hydrogen's one-point lead over Argon on the Intelligence Index is not a lead over the market. It is a lead over Google's own previous model at one setting.
What to watch
Whether developers use Hydrogen as a default or a fallback will show up in Orchestrator behaviour first. Google's pricing makes the model attractive for high-volume agent work, but its value depends on how often it finishes a task without escalation. If Orchestrator ends up calling Carbon for most steps, the savings shrink.
Independent cost-per-task data is the next check. Artificial Analysis has a number, and Vals has a number, but both are benchmark estimates. Real invoices from teams running coding agents will show whether the flat rate survives contact with retries, tool calls and long context.
The free tier is the quieter variable. If automatic model selection sends most free prompts to Hydrogen, Google gets a large distribution win and a higher inference bill. If it keeps free users on lighter models and reserves Hydrogen for paid tiers, the launch is narrower than the language implies.
Google has shipped the model, named the price and published a safety case. Whether a cheaper model that is almost as good is enough to move a market that has spent two years paying for the most capable one is a question for the next invoice.
Cover image: official Google key art for Gemini 4 Hydrogen, cropped to the crystal, courtesy of Google.
Tags
Frequently asked questions
- How much does Gemini 4 Hydrogen cost?
- Google charges $1.50 per million input tokens and $7.50 per million output tokens, with the same rate across the model's 1 million-token context window. Unlike Gemini 4 Carbon, Hydrogen has no introductory discount that expires on 30 November.
- Is Gemini 4 Hydrogen better than Gemini 4 Argon?
- On Artificial Analysis's Intelligence Index, Hydrogen scores 51 at its default high thinking level and 54 at xhigh, against 50 for Argon. Google positions Carbon above both for the hardest reasoning. The trade-off is that Hydrogen is tuned for speed and cost, not for long-horizon reasoning.
- Where can I use Gemini 4 Hydrogen?
- Developers can use it in Google AI Studio, the Gemini API and Android Studio. Enterprises get it through Gemini for business and Antigravity. Consumers get it in the Gemini app and Google Search, with free app users reaching it through automatic model selection.
- Does Gemini 4 Hydrogen have a 1 million-token context window?
- Yes. Google lists 1 million tokens of input and up to 1 million tokens of output, with four thinking levels: low, medium, high as the default, and xhigh.








