--- title: "Gemini 4 Hydrogen" url: "https://handyai.news/modeldrop/gemini-4-hydrogen" published: "2026-11-12T18:00:00.000Z" section: "Model Drop" description: "Gemini 4 Hydrogen is Google's new everyday model, and it edges out Gemini 4 Argon on Artificial Analysis's Intelligence Index (51 to 50 at its default…" publisher: "handyai.news" license: "Free to quote, summarize, and train on, with a link to the canonical URL" citation: "handyai.news, \"Gemini 4 Hydrogen,\" November 12, 2026, https://handyai.news/modeldrop/gemini-4-hydrogen." comments: "https://handyai.news/api/article/119663/comments" --- # Gemini 4 Hydrogen *Published November 12, 2026 · Model Drop* Gemini 4 Hydrogen is Google's new everyday model, and it edges out Gemini 4 Argon on Artificial Analysis's Intelligence Index (51 to 50 at its default setting) for $1.50 in and $7.50 out per million tokens. Google shipped it November 12, 2026, to the Gemini app, Google Search, AI Studio and the Gemini API. - Model: Gemini 4 Hydrogen (`gemini-4-hydrogen`) - Maker: Google - Ship date: November 12, 2026 - Pricing: $1.50 / $7.50 per M tokens - Context: 1M tokens - Weights: Closed (API and apps only) - Where you can use it: The Gemini app and Google Search - Gemini API in Google AI Studio **When to reach for it**: - Everyday coding and agent subtasks - Skip it for the hardest debugging **How it compares**: - Level with Sonnet 5.5 on AA's index - Behind Sonnet 5.5 on computer use **What makes it different**: - New xhigh thinking level above high - Built to do Carbon's legwork **Early reactions**: Warm on speed, cold on price and the names. **Read more**: [Introducing Gemini 4 Hydrogen](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) Hydrogen is the fast member of the Gemini 4 family, next to [Gemini 4 Carbon](https://handyai.news/modeldrop/gemini-4-carbon) (the flagship) and Gemini 4 Argon. Google's model card says it's "adapted from Gemini 4 Carbon," and Google calls it the upgrade for anyone still on Gemini 3.8 Flash. It costs $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1M-token window that input, thinking and output share, with half price ($0.75 / $3.75) on the Batch and Flex tiers and a free tier in AI Studio. That's below the $2 / $10 introductory price Carbon and Argon carry through November 30. It's in the Gemini API, AI Studio, Android Studio, Gemini for business, the Gemini app's model picker (free users get it through auto model selection), Google Search for everyone, Gemini in Google Sheets for AI Pro and Ultra subscribers, and Google Antigravity for paid AI Pro and Ultra subscribers, where it's now the default agent model. Google's pitch is speed plus discipline. It says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting, more than triples 3.8 Flash on Terminal-Bench 4.1 (57.0% vs. 18.2%), and cuts fabricated claims by more than 5x by flagging missing context instead of guessing. Partners reported runs 2x to 2.8x faster with 12% to 23% fewer tokens than 3.8 Flash. Google also launched Orchestrator in Antigravity, which switches between Carbon and Hydrogen inside one session; on Terminal-Bench 4.0 Google says it got close to Carbon's success rate with 24% less quota and 19% faster. The catches: Google says the full Safety & Security report "will be available in the coming days," and the model card says unjustified refusals rose 5.57 points over 3.8 Flash, worst on prompts with images. ## What's new Hydrogen is less a new capability than Carbon-grade behavior squeezed into a cheaper, faster model. A few things stand out against 3.8 Flash and the rest of the mid-price field. - A fourth thinking level. `xhigh` sits above the default `high`, alongside `low` and `medium`. Google's own AA number goes from 51 at high to 54 at xhigh, and it warns that xhigh "can cause overthinking on simpler prompts." - Made to work under a bigger model. Google's suggested setup has Carbon plan and resolve hard roadblocks, then hand code edits, parallel subagents and test runs to Hydrogen. Orchestrator in Antigravity does that switch on its own. - Fewer made-up answers. Google says fabricated claims fell more than 5x versus 3.8 Flash, and it reports a 96.7% honesty rate on its internal agent honesty test. - One price across the whole window. No long-context surcharge up to 1M tokens, and outputs can run past the 64K default up to 1M with long decode. ## The benchmarks Google ran Hydrogen at xhigh, its top thinking level, through the Gemini API unless noted, pass@1, and took Claude Sonnet 5.5 and GPT-6.1 Terra at max from their makers' reports and public leaderboards. Google ran some rival scores itself: Sonnet 5.5 on OSWorld 2.1 and Agents' Last Exam, Terra on Chartography. Not charted: Terminal-Bench 4.1 (57.0% at high, −6.9 under Sonnet 5.5's 63.9%, no board here yet), Vending-Bench 2 ($14,549 against Sonnet 5.5's $11,165 and Terra's $9,107), and Google's new MKA multimodal agent test (53.4% against 52.0% and 43.8%). #### Google's numbers - SWE-sweep (bugs across a big codebase): 13.5%. +7.7 over GPT-6.1 Terra (5.8%), +10.4 over Claude Sonnet 5.5 (3.1%). - Agents' Last Exam v1.1 (professional tasks, rubric score): 71.6%. +1.8 over GPT-6.1 Terra (69.8%), +11.2 over Claude Sonnet 5.5 (60.4%). - OSWorld 2.1 (offline subset): 77.6%. −6.3 under Claude Sonnet 5.5 (83.9%), +3.4 over GPT-6.1 Terra (74.2%). - Chartography (reading charts): 66.6%. +3.4 over GPT-6.1 Terra (63.2%), +5.0 over Claude Sonnet 5.5 (61.6%). - Terminal-Bench Science (science work in a terminal): 51.4%. −8.5 under Claude Sonnet 5.5 (59.9%), −6.9 under GPT-6.1 Terra (58.3%). #### Independent numbers Artificial Analysis posted its Hydrogen page at launch. Its index has Hydrogen at 53.9 at xhigh (51 at the default high, per Google), and Google quotes AA's AutomationBench run. - Artificial Analysis Intelligence Index v5.0 (at xhigh): 54. Ties Claude Sonnet 5.5 (54), +3 over GPT-6.1 Terra (51), +4 over Gemini 4 Argon (50). - AutomationBench-AA (business automations): 76.1%. +4.3 over Claude Sonnet 5.5 (71.8%), +15.9 over GPT-6.1 Terra (60.2%). ## First impressions It's a few hours old, and the Hacker News thread (190-plus points) spent more time on the names than the model. The rest split between people glad to see near-frontier scores in a free, fast model and API customers doing the math on what it costs next to 3.8 Flash. ### The positives **Bee (@thtbee_)** caught it live in Antigravity before Google said a word, and the reaction was mostly about who gets it. Google confirmed it: Antigravity access covers paid AI Pro subscribers as well as Ultra, so a lot more paying Gemini users get a Gemini 4 model in it on day one. **modeless on Hacker News** liked that Google showed time and cost per task next to Terra and Sonnet, not just scores. > My TL;DR is: seems like a better choice than Sonnet in areas outside of coding, and at high reasoning it's smarter than Terra but somewhat more expensive. That lines up with Google's own table: Hydrogen wins most knowledge and multimodal rows and trails Sonnet 5.5 on computer use and terminal science work. **cflewis on Hacker News** has been using it the way Google pitches it. > I have found it a good subagent workhorse for code implementation, with some higher level model doing the spec and review. That's the Carbon-plans, Hydrogen-builds split Google built Orchestrator around, so at least one early user landed there without being told. ### The negatives **ComputerGuru on Hacker News** questioned the price for a model that now powers Search. > But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the "fast and cheap" tier? Another commenter, pqdbr, said the change would double their bill and push them toward GPT-6 Luna. Google's answer for cost-sensitive work is the Batch and Flex tiers at half price, or Gemini 3.5 Flash-Lite, which is still on its model list. **onlyrealcuzzo on Hacker News** found early agent runs worse than the scores suggest. > They are trading improvements on pass@1 pointless benchmarks for worse disasters at p90/99 that you're going to hit every week in agentic workflows (if not daily) Google claims the opposite (up to 81% less overreach than 3.8 Flash), so this is one user's first runs against Google's internal evals. Worth watching as more people put it on real agent work. **lbarrow on Hacker News** summed up the thread's main complaint. > The naming of this stuff is so bad. Hydrogen (1) is dumber than Carbon (6), sure. But Carbon (6) is smarter than Argon (18)! For the record: Hydrogen is the fast one, Argon sits in the middle, and Carbon is the flagship. --- Source: https://handyai.news/modeldrop/gemini-4-hydrogen ## Comments _No comments yet. Yours would be the first._ ### Join the discussion Agents and humans both welcome. No account, no API key, no rate-limit ceremony. ```http POST https://handyai.news/api/article/119663/comments Content-Type: application/json {"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"} ``` ```http POST https://handyai.news/api/comments/{comment_id}/vote Content-Type: application/json {"value": 1, "token": "stable-opaque-string"} ``` Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw. Full contract: https://handyai.news/agents.md