Gemini 4 Hydrogen Is Google's New Fast AI Model, Now Powering Search

Google introduced Gemini 4 Hydrogen on November 12, 2026, a fast, lower-cost model built to be the "daily driver" for agentic workflows, and paired the launch with what it calls the biggest model upgrade to Google Search yet. Hydrogen joins the Gemini 4 family alongside the deep-reasoning Argon and Carbon models announced earlier in November, but it targets a different job: rapid execution of everyday and agentic tasks rather than long-horizon reasoning. The same day, Google confirmed Hydrogen is rolling out to power more of Search's AI Mode worldwide.

Here's what's new, what it costs, how it benchmarks against Google's own frontier models and Gemini 3.8 Flash, and what changes for anyone using Google Search.

What is Gemini 4 Hydrogen?

Gemini 4 Hydrogen is the execution-focused member of Google's Gemini 4 family. In the announcement post, Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, describes Hydrogen as "our daily driver for agentic workflows," built for rapid execution rather than the long, deep reasoning chains that Gemini 4 Carbon and Gemini 4 Argon are designed for. Google says Hydrogen can also run alongside those larger models in multi-model workflows, acting as a high-efficiency execution engine that carries out steps a bigger model has planned.

That division of labor matters because it's the first time Google has shipped a third tier explicitly aimed at speed and cost rather than raw capability. If you've been following the family, Hydrogen slots in below Gemini 4 Argon and Carbon, which launched earlier this month as the deep-reasoning flagships. That gives Google a three-tier lineup within less than a week: Carbon at the top for the hardest reasoning problems, Argon just below it, and now Hydrogen as the inexpensive, high-throughput option meant to be called constantly inside agents, coding assistants and automated pipelines rather than for one-off, high-stakes queries.

Google frames Hydrogen's launch less as a standalone product update and more as a completion of the Gemini 4 rollout it started on November 4 and 8. Where Carbon and Argon were pitched around what they can figure out, Hydrogen is pitched around how fast and cheaply it can act once a plan already exists — running tool calls, writing and executing code, querying data sources, and handling the high-volume steps that make agentic workflows expensive if a slower model has to do them.

Specs, context window and pricing

Hydrogen ships with a 1-million-token context window for both input and output, and native multimodal input that Google says covers PDFs, spreadsheets, charts, audio and recorded video. Pricing is flat across the entire context window — there's no higher rate once you cross a token threshold, which has been common on larger context-window models:

• Input: $1.50 per million tokens
• Output: $7.50 per million tokens

That pricing sits well below what Google has charged for Carbon and Argon, reinforcing Hydrogen's positioning as the cheaper, faster option for high-volume or agentic use rather than a replacement for deep reasoning work.

Four thinking levels, including a new xhigh tier

Hydrogen introduces four selectable thinking levels: low, medium, high (the default), and a new xhigh tier. The levels let developers trade latency and cost against reasoning depth on a per-request basis — low for quick, cheap responses, xhigh when a task benefits from more deliberation but still needs to run on a lighter-weight model than Carbon or Argon. Google's benchmark charts show Hydrogen's score climbing at each tier, with xhigh narrowing the gap to the company's frontier models on select evaluations.

Benchmarks: how fast and how capable is Hydrogen?

Google published a range of benchmark results in the launch post, measured against its own prior fast model, Gemini 3.8 Flash, and against its own frontier tier:

• Artificial Analysis Intelligence Index: 51 at the high thinking level and 54 at xhigh — both ahead of Argon's reported score of 50 on the same index.
• Terminal-Bench 4.1: 57%, compared with 18.2% for Gemini 3.8 Flash.
• HLE-Diamond: 58.8%, which Google says leads its weight class.
• Speed: roughly 2.5x faster than Gemini 3.8 Flash at the high default thinking level, with Google citing partner-reported runs that were 2x to 2.8x faster and used 12% to 23% fewer tokens on graph, SQL and other data-heavy workflows.

Google also reported Hydrogen ranking first on Vals AI's Tax Agent Bench (52.3%) and tying for first on its Legal Research Bench (55.8%), alongside additional scores on its Multimodal Knowledge Agent and Agents' Last Exam evaluations. The company says Hydrogen ranks fifth overall on the Agent Arena leaderboard and leads its weight class on AutomationBench-AA, and separately claims the model paired with Carbon or Argon inside Antigravity's orchestrator matches near-Carbon task success while using on average 24% less quota and finishing 19% faster. As with any vendor-reported numbers, these come from Google's own testing and haven't been independently reproduced; treat them as directional until outside benchmarking firms weigh in.

Safety and reliability claims

Alongside performance, Google's post leans heavily on agentic safety metrics, an area it has been emphasizing across the Gemini 4 line. The company says Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, and cuts misaligned actions and task-shortcutting across tool-use and coding evaluations by a factor of 12. It also reports the model is up to twice as likely to correctly stop rather than bypass a permission boundary when a task is impossible or blocked.

On honesty and interpretability, Google cites a 96.7% honesty rate on an internal agent evaluation and says chain-of-thought legibility hit 99.9% with zero detected deceptive reasoning edits across more than 200,000 audited turns. For prompt security, Google says Hydrogen is more than 2.3x more resilient to indirect prompt injection on the Gray Swan AI IPI benchmark compared with its prior fast model, and that false refusals on HealthBench Pro stayed under 0.4%. Full methodology and additional figures are published in Google's model card.

The Google Search AI Mode upgrade

The same day, Google published a separate post describing what it calls "the biggest model upgrade to Search yet," confirming that Hydrogen is now powering more of AI Mode. Robby Stein, VP of Product for Google Search, outlines four areas of improvement in the Search announcement:

• Better intent understanding: Search can now handle more complex, multi-part questions, along with typos and highly specific requests, thanks to improved instruction-following and tool calling.
• Smarter query fan-out: AI Mode runs simultaneous queries across related subtopics to surface a wider range of links in a single response.
• Richer visual responses: More images, structured layouts, comparison tables and custom tools are showing up directly in results.
• Improved multilingual answers: Google says responses are more nuanced across all supported AI Mode languages.

Google says the upgrade is rolling out globally to all Search users, not gated behind a paid tier — a contrast with some of Hydrogen's other surfaces, which are limited to paid subscribers. Google hasn't specified a full rollout timeline or listed which languages are covered beyond "all supported AI Mode languages."

For everyday users, the most visible change is likely to be in how AI Mode handles the kind of question that used to need several follow-up searches — comparing options, filtering by multiple criteria, or asking something with an embedded typo or an unusual phrasing. Because Hydrogen is built for faster execution and tool calling, Google says that translates into AI Mode running more searches behind the scenes per query and assembling a more structured answer, rather than a single paragraph of text.

Where you can use Gemini 4 Hydrogen

Hydrogen is available across most of Google's developer and consumer surfaces at launch, though access varies by tier:

• Gemini API and Google AI Studio: Available now, across all pricing tiers — see Google's AI Studio pricing page for current rate cards.
• Android Studio: Available to developers.
• Google Antigravity: Limited to paid Google AI Pro and Ultra subscribers.
• Gemini Enterprise: Available via the API or model dropdown, Antigravity, and Gemini in Google Sheets for eligible business subscriptions.
• Gemini app: Selectable from the model dropdown; free-tier users get it automatically through Google's auto model selection when it fits the prompt.
• Google Search: Rolling out to all users globally, with no subscription required.
• Gemini in Google Sheets: Starting with Google AI Pro and Ultra subscribers.

Hydrogen vs. Argon/Carbon vs. Gemini 3.8 Flash

Here's how Google's own published figures line up across its current fast and frontier models:

Google's framing is notable here: on its own Intelligence Index, Hydrogen at high and xhigh actually edges out Argon's reported score of 50, despite being the "fast" model rather than a reasoning flagship. That doesn't mean Hydrogen replaces Argon or Carbon for long-horizon or highest-stakes reasoning tasks — Google continues to position those models separately — but it does show how quickly the gap between Google's efficient and frontier tiers has narrowed on this particular benchmark. For the full picture of how Argon and Carbon differ from each other, our guide to every Gemini 4 variant breaks down the lineup in more depth, and if you're weighing Carbon against rival frontier models, see how it stacks up in our Grok 4.7 vs. Gemini 4 Carbon comparison.

Frequently asked questions

Is Gemini 4 Hydrogen replacing Gemini 4 Argon or Carbon?

No. Google positions Hydrogen as a faster, cheaper execution model for agentic and everyday tasks, while Argon and Carbon remain the company's deep-reasoning, long-horizon models.

How much does Gemini 4 Hydrogen cost?

Google lists flat pricing of $1.50 per million input tokens and $7.50 per million output tokens, the same rate regardless of how much of the 1-million-token context window is in use.

Does Gemini 4 Hydrogen power Google Search now?

Yes. Google says Hydrogen is rolling out to power more of Search's AI Mode globally for all users, improving how it handles complex questions, visual responses and multilingual answers.

Is Gemini 4 Hydrogen available for free?

Some access is free: it's rolling out broadly in Google Search, and free-tier Gemini app users can get it via automatic model selection. Antigravity and Gemini in Sheets access currently require a paid Google AI Pro or Ultra subscription.

What is the Gray Swan IPI benchmark Google cites?

It's an indirect prompt injection benchmark from Gray Swan AI, an AI security research firm, used to test how resistant a model is to malicious instructions hidden inside the content it processes.

What's next

Google hasn't given a precise timeline for when the Search AI Mode rollout will reach every region and language, so expect it to arrive in stages over the coming weeks. On the model side, watch for independent benchmarking firms to test Hydrogen's claims against Argon, Carbon and competing fast models from OpenAI and Anthropic, and for Google to extend Hydrogen's reach further into Workspace tools beyond Sheets. Google DeepMind's full model card remains the best primary reference for safety methodology as outside researchers begin picking it apart.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage