Tag: Gemini 4 Hydrogen

  • Google Launches Gemini 4 Hydrogen, Its Fast Everyday Gemini 4 Model (2026)

    Google Launches Gemini 4 Hydrogen, Its Fast Everyday Gemini 4 Model (2026)

    Takeaway: On November 12, 2026, Google introduced Gemini 4 Hydrogen, a Gemini 4 model built for fast everyday coding and agentic work rather than the long-horizon reasoning of Gemini 4 Carbon and Gemini 4 Argon. The model ID is gemini-4-hydrogen. Google’s pricing page lists a standard paid price of $1.50 per 1 million input tokens and $7.50 per 1 million output tokens, including thinking tokens, billed flat across a unified 1,048,576-token context window. As of November 16, 2026, that official standard paid price was unchanged. Hydrogen is rolling out in the Gemini API and Google AI Studio, the Gemini app, Google Search, Gemini for business, Gemini in Google Sheets for eligible subscriptions, and Google Antigravity for paid Google AI Pro and Ultra subscribers. This is a separate launch from this week’s Gemini Decisions API and Decision 1.

    The official Gemini 4 Hydrogen model page and the What’s new in Gemini 4 Hydrogen guide are the right starting points for developers. The DeepMind model card, also published November 12, 2026, adds limits Google’s launch post does not dwell on, including a March 2026 knowledge cutoff and a Safety & Security Report that Google says will follow in the coming days.

    Where Hydrogen sits next to Carbon and Argon

    Google’s framing is a split inside the Gemini 4 family. Carbon and Argon are the frontier models for deep, long-horizon work. Hydrogen is the fast execution model: writing code, running tests, verifying changes, and handling everyday assistant tasks. Koray Kavukcuoglu, SVP at Google DeepMind and Google’s chief AI architect, is the byline on the main announcement. A companion developer post repeats the same positioning for repository coding, interactive web development, and multimodal workflows.

    Hydrogen can stand on its own, or it can sit beside Carbon and Argon as the cheaper, faster worker in a multi-model loop. That is also the idea behind Orchestrator, a new Google Antigravity option described below. It is not a drop-in replacement for Carbon, and it is not a decision-scoring endpoint like Gemini Decision 1. If you need long-horizon architectural judgment, you still want a frontier model. If you need a closed-set probability, you still want the Decisions API.

    Official price and API limits

    The November 12 launch post says Hydrogen is “priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.” The live Gemini API pricing page matches those standard paid figures and adds the rest of the rate card. Confirm that page before you budget, because Google notes that discounts and availability can change.

    • Standard paid: $1.50 per 1 million input tokens and $7.50 per 1 million output tokens, including thinking tokens. Implicit context caching is $0.075 per 1 million tokens. Explicit caching is not supported.
    • Free tier: input and output, including thinking tokens, are listed as free of charge. Implicit caching is also listed as free of charge. Google’s terms row says free-tier usage may be used to improve its products; paid standard, Batch, Flex, and Priority usage is listed as not used that way.
    • Batch and Flex: $0.75 input and $3.75 output per 1 million tokens, with implicit caching at $0.0375. No free-tier row is listed for those consumption options.
    • Priority: $2.70 input and $13.50 output per 1 million tokens, with implicit caching at $0.135. A free-tier row is listed as free of charge.
    • Grounding: on the paid standard and Priority tiers, Grounding with Google Search and Grounding with Google Maps are listed with a shared monthly allotment of 5,000 free requests across Gemini 3 and newer models, then $14 per 1,000 requests. Search grounding is also subject to a daily cap documented on the tools pricing section. Grounding is listed as not available on the free tier. A customer request can trigger more than one Search query, and Google says you are charged for each query.

    The model page lists a unified context window of 1,048,576 tokens, with a default synchronous output limit of 65,536 tokens and up to 1,048,576 tokens with long decode. Inputs are text, images, video, audio, and PDF. Output is text. Thinking levels are low, medium, high, and a new xhigh tier; the launch post says high is the default. Supported capabilities include code execution, computer use, function calling, structured outputs, file search, URL context, and implicit context caching. Audio generation, image generation, and the Live API are listed as not supported. The stable model code is gemini-4-hydrogen.

    The model card describes Hydrogen as adapted from Gemini 4 Carbon. It also warns of general limits such as hallucinations, occasional slowness or timeouts, and higher token use at higher thinking levels, and it sets the knowledge cutoff at March 2026, “though in some domains the model’s knowledge may be less up to date.” Those are official limits, not reviewer caveats.

    What Google claims on speed and quality

    The launch post compares Hydrogen mainly with Gemini 3.8 Flash, which teams would be leaving if they migrate, and with Gemini 4 Argon on one headline index. These are Google’s figures. We have not rerun the harnesses.

    • On the Artificial Analysis Intelligence Index, Google says Hydrogen scores 51 at its default high thinking level, ahead of Argon’s 50, and 54 at the xhigh setting, “while often completing tasks faster than models in its class.”
    • As a standalone upgrade from 3.8 Flash, Google says Hydrogen finishes tasks 2.5 times faster at its high default setting, more than triples Terminal-Bench 4.1 scores (57% vs. 18.2%), cuts fabricated claims by more than 5x, and “adapts cleanly when a tool call fails or a search returns empty.”
    • For repository work, Google says partners have seen Hydrogen run 2x to 2.8x faster than 3.8 Flash in early testing across complex graphs, SQL, and data-heavy enterprise workflows, while reducing token usage by 12% to 23%. That is a partner-testing claim, not a public leaderboard.
    • Google also cites class or overall leads on several other tests: HLE-Diamond at 58.8%, Vals Tax Agent Bench at 52.3% (ranked #1 overall), Vals Legal Research Bench at 55.8% (tied for #1), a class lead on AutomationBench-AA, Multimodal Knowledge Agent (MKA) at 53.4%, and Agents’ Last Exam v1.1 at 71.6%. On Agent Arena, arena.ai’s live benchmark, Google says Hydrogen ranks #5 for multi-step tasks such as deep research, data analysis, and broad codebase navigation.

    The official model card publishes a smaller table, dated November 2026, that is worth reading next to those marketing comparisons. On Terminal-Bench 4.1 (cheating-adjusted), the card lists Hydrogen at 57.0% and Claude Sonnet 5.5 at 63.9%. On OSWorld-2.1 (offline subset), it lists Hydrogen at 77.6% and Sonnet 5.5 at 83.9%. Hydrogen leads the same table on SWE-sweep (13.5% vs. 3.1% for Sonnet 5.5 and 5.8% for GPT-6.1 Terra), Agents’ Last Exam v1.1, AutomationBench-AA, Vending-Bench 2, MKA, Chartography, and HLE-Diamond. In other words, Google’s “more than triples 3.8 Flash” line on Terminal-Bench 4.1 is consistent with the card, but the same official table does not show Hydrogen leading every comparable model on every coding or computer-use test. Treat the launch post as positioning, and the model card as the fuller score sheet.

    Orchestrator in Google Antigravity

    Alongside Hydrogen, Google is introducing Orchestrator in Google Antigravity. Instead of pinning a whole session to one model, Orchestrator switches between models such as Carbon and Hydrogen as a task evolves. Google’s example is Carbon exploring a repo and writing a plan, Hydrogen making edits and running tests, and Carbon taking the session back if Hydrogen hits friction.

    The main launch post says that on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster. The Antigravity post reports the same 24% and 19% figures on Terminal-Bench 4.1. Those are the same percentages on two differently numbered benchmark references, so we are leaving both citations as Google wrote them rather than collapsing them into one bench version.

    Availability is narrower than Hydrogen itself. The Antigravity post says Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers. Orchestrator (Carbon) and Orchestrator (Argon) are rolling out for Google AI Ultra subscribers, and Orchestrator (Argon) for paid Google AI Pro subscribers, with more models and plans “supported in the future.” Enterprises can access both in Gemini for business. For teams that want to wire the models themselves, Google also points to explicit multi-model pipelines in the Antigravity SDK, with Hydrogen handling higher-volume triage and UI iteration and Carbon reserved for larger refactors, new features, and security checks. That SDK example is Google’s demo, not a required architecture.

    Gemini app, Google Search, and Sheets

    Hydrogen is not only a developer model. Robby Stein, VP of product for Google Search, calls it “the biggest model upgrade to Search yet.” Google says Hydrogen is available globally in Search for all users, with better handling of multi-part questions, typos, and highly specific requests; more visual responses, including images, comparison tables, and structured layouts; richer responses across supported AI Mode languages; and an upgraded query fan-out that Google says connects people to a broader range of links. Those are Google’s product claims about Search quality. They are not a traffic or ranking guarantee for any particular site.

    In the Gemini app, the launch post says Hydrogen follows tone, format, and rules more closely, calls tools and skills more often, and can pull context from uploads, Google apps such as Gmail and Drive, and Connected Apps such as Airtable and Dropbox. Free users get Hydrogen through auto model selection “when it’s the appropriate model for the prompt.” Everyone else can pick it from the model drop-down, according to Google. Gemini in Google Sheets starts with Google AI Pro and Ultra subscribers, and with eligible business subscriptions. Google says Hydrogen can generate forecasting models in Sheets canvas more than 30% faster than previous models; that is a company figure.

    Safety claims — and what is still forthcoming

    Google says Hydrogen inherits the Gemini 4 safety and security stack and improves on 3.8 Flash in agentic settings: up to 81% less overreach, up to 79% less under-service, 12x fewer misaligned actions and task shortcuts on tool-use and coding evaluations, up to 2x less likely to bypass permission boundaries on impossible or blocked tasks, a 96.7% honesty rate on an internal agent honesty evaluation, 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns, and more than 2.3x better resilience on the Gray Swan indirect prompt injection benchmark. It also says updated safeguards keep false refusals on HealthBench Pro under 0.4%. All of those figures are Google’s.

    The model card is more cautious in places. The full Safety & Security Report “will be available in the coming days.” Automated development evaluations versus 3.8 Flash show small gains on text and multilingual safety, a larger gain on image-to-text safety (+7.00 percentage points), a slight regression on refusal tone (−0.65pp), and more unjustified refusals (+5.57pp), “particularly for multimodal prompts.” Google says a manual review found those losses were overwhelmingly false positives or not egregious, and that the evaluation methods changed, so the numbers are not directly comparable with older Gemini model cards.

    On Frontier Safety, Google says a Material Capability Change assessment did not find a material increase versus Gemini 4 Argon, so a full critical-capability suite was not required. Targeted CBRN and cyber tests found that Hydrogen does not reach those critical capability levels, though it reaches the alert thresholds for both. Mitigations continue. Readers who need the formal assessment should wait for the forthcoming report rather than treat the launch-post percentages as the complete safety record.

    Availability checklist

    Google’s “available to everyone today” line still has surface-by-surface limits. This is the rollout as the November 12 posts describe it:

    • Developers: Gemini API in Google AI Studio and Android Studio. The stable ID is gemini-4-hydrogen.
    • Google Antigravity: paid Google AI Pro and Ultra subscribers, with Orchestrator gated as described above.
    • Enterprises: Gemini for business, via the API or the model drop-down, plus Antigravity and Gemini in Google Sheets for eligible business subscriptions.
    • Gemini app: model drop-down for users who have it; free users via auto model selection when Google decides Hydrogen is the right model.
    • Google Search: global rollout for all users.
    • Gemini in Google Sheets: starting with Google AI Pro and Ultra subscribers.

    Gemini 4 Hydrogen facts at a glance

    ItemWhat Google’s November 12 announcement, API docs, and model card say
    ModelGemini 4 Hydrogen, model ID gemini-4-hydrogen; adapted from Gemini 4 Carbon
    AnnouncedNovember 12, 2026, by Koray Kavukcuoglu, Google DeepMind
    RoleFast everyday coding and agentic execution; not a replacement for Carbon, Argon, or the Decisions API
    Context / outputUnified 1,048,576-token window; default synchronous output 65,536 tokens; up to 1,048,576 with long decode
    Thinking levelslow, medium, high (default, per the launch post), and xhigh
    Standard paid price$1.50 / $7.50 per 1 million input / output tokens, including thinking tokens; implicit cache $0.075; no explicit cache
    Other paid tiersBatch and Flex $0.75 / $3.75; Priority $2.70 / $13.50
    Free tierInput and output listed as free of charge on the standard (and Priority) rows; data-use terms differ from paid
    Headline Google comparisonsAA Index 51 default / 54 xhigh vs. Argon 50; Terminal-Bench 4.1 57% vs. 3.8 Flash 18.2%; 2.5x faster than 3.8 Flash at high default
    Model-card comparison worth notingTerminal-Bench 4.1: 57.0% vs. Claude Sonnet 5.5 at 63.9%; knowledge cutoff March 2026
    OrchestratorAntigravity option that switches between models such as Carbon and Hydrogen; Google cites 24% less quota and 19% faster, on Terminal-Bench 4.0 in the main post and 4.1 in the Antigravity post
    Not supported in the API docsAudio generation, image generation, Live API, and explicit context caching

    FAQ

    What is Gemini 4 Hydrogen?

    It is a Gemini 4 model Google announced on November 12, 2026, for fast everyday coding, agentic workflows, the Gemini app, and Google Search. The API model ID is gemini-4-hydrogen. Google positions it as the rapid-execution counterpart to Gemini 4 Carbon and Gemini 4 Argon, which remain the frontier models for harder, longer-horizon work.

    How much does Gemini 4 Hydrogen cost?

    On the standard paid Gemini API tier, Google lists $1.50 per 1 million input tokens and $7.50 per 1 million output tokens, including thinking tokens, with the same rate across the 1-million-token context window. Batch and Flex are listed at $0.75 / $3.75, and Priority at $2.70 / $13.50. A free tier lists input and output as free of charge. Confirm the live pricing page for grounding, caching, and data-use terms.

    Is Gemini 4 Hydrogen better than Gemini 4 Carbon?

    Not as a general replacement. Google built Carbon and Argon for deep, long-horizon reasoning and built Hydrogen for speed and everyday execution. Orchestrator exists because Google expects teams to use both. On some public-facing comparisons Hydrogen even edges Argon on the Artificial Analysis Intelligence Index at default and xhigh settings, but that is not the same as outperforming Carbon on the hardest tasks.

    How is this different from Gemini 3.8 Flash?

    Google is presenting Hydrogen as the successor for teams that used 3.8 Flash as their high-volume workhorse. The launch post’s head-to-head claims versus Flash include 2.5x faster task completion at the high default setting, a Terminal-Bench 4.1 jump from 18.2% to 57%, and more than 5x fewer fabricated claims. The model name no longer uses the Flash label.

    Can I use Gemini 4 Hydrogen in Google Search and the Gemini app?

    Yes, according to Google. Hydrogen is rolling out globally in Search for all users. In the Gemini app, it appears in the model drop-down, and free users can get it through auto model selection when Google considers it the right model for the prompt.

    What is Orchestrator?

    It is a Google Antigravity model option, announced with Hydrogen, that dynamically switches between models such as Carbon and Hydrogen in one session. It is rolling out first to paid Google AI Ultra subscribers as Orchestrator (Carbon) and Orchestrator (Argon), and to paid Google AI Pro subscribers as Orchestrator (Argon). Google’s efficiency figures for it should be read as company benchmark results.

    Does Hydrogen support a 1 million token output?

    The API docs list a unified 1,048,576-token context window. Default synchronous output is 65,536 tokens. Outputs up to the full 1,048,576-token window require long decode. The DeepMind model card describes the same idea as “up to 64K default synchronous output tokens / up to 1M via background long decode.”

    Should I treat the benchmark numbers as independent results?

    No. The speed, quality, and safety percentages in the launch post are Google’s. The model card adds an official comparison table, including tests where Claude Sonnet 5.5 still leads, and it flags a March 2026 knowledge cutoff plus a safety report that has not been published yet. The useful test is still your own workloads, latency budget, and thinking-level settings.