Gemini 4 Hydrogen: What It Is, What It Changes, and Who Should Care
Quick answer: Gemini 4 Hydrogen is Google’s fast, lower-cost Gemini 4 model, launched on November 12, 2026. It is built for everyday coding and multi-step agent work, it replaces Gemini 3.8 Flash as the default workhorse, and it is the first Gemini 4 model most people will actually meet — because it now powers the hardest questions in Google Search’s AI Mode and is rolling out across the Gemini app, Google AI Studio, and the Gemini API. It is not Google’s smartest model. It is the fast one, and Google priced it to be used constantly: $1.50 per million input tokens and $7.50 per million output tokens.
I spent this post the way I approach every model launch: reading Google’s own announcement and model card first, then checking every headline claim against a second source before repeating it. I did not run a live signed-in session in AI Studio for this one — the model had been out for six days when I wrote this, and the verifiable story is in Google’s own documents, pricing page, and the independent benchmark tables — so this post says exactly what I verified, and flags where Google’s numbers are Google’s numbers.
The short version: if you use Google Search or the Gemini app, Hydrogen is already changing your answers. If you build with the Gemini API, it is the new default to test before reaching for the bigger models. And if you are choosing between models, the interesting question is no longer “which model is best” — it is which model should do which part of the job.
What is Gemini 4 Hydrogen?
Gemini 4 Hydrogen is the third member of Google’s Gemini 4 family, and the one with the clearest job description. While Gemini 4 Carbon and Gemini 4 Argon are the frontier models built for deep, long-horizon reasoning on hard problems, Hydrogen is built for rapid execution across everyday coding and agentic workflows: writing code, running tests, verifying changes, doing research, and carrying multi-step tasks to completion. Google calls it its “daily driver for agentic workflows.”
It is adapted from Gemini 4 Carbon, according to the official Gemini 4 Hydrogen model card, and it accepts text, images, audio, video, and PDF input, with text output. Its context window is 1,048,576 tokens, shared by input, thinking, and output, with up to 65,536 tokens per synchronous response and up to 1 million tokens via background long decode.
The name is the tell. Hydrogen is the light, fast element in the lineup — Carbon and Argon do the heavy reasoning; Hydrogen does the running. That division of labor is the whole point of this release, and it is also why Google launched a second thing alongside the model: Orchestrator, a feature in Google Antigravity that switches between models inside a single session, letting Carbon plan while Hydrogen executes. More on that below.
How does Hydrogen fit into the Gemini 4 lineup?
Google now sells three Gemini 4 models, and the differences are about role, not just size. Here is the lineup as Google describes it in the launch announcement:
| Model | Built for | Context window | API price (per 1M tokens) | Free tier |
|---|---|---|---|---|
| Gemini 4 Carbon | Deep, long-horizon reasoning; hardest problems | 1M tokens | $2 in / $10 out through Nov 30, then $4 / $20 | No |
| Gemini 4 Argon | Frontier reasoning; architecture and complex analysis | 1M tokens | $2 in / $10 out through Nov 30, then $4 / $20 | No |
| Gemini 4 Hydrogen | Fast execution; everyday coding and agentic work | 1M tokens (shared input + thinking + output) | $1.50 in / $7.50 out, flat | Yes — API, AI Studio, Gemini app, Search |
Two things in that table matter more than the rest. First, Hydrogen is the only Gemini 4 model with a free tier on the Gemini API — Carbon and Argon are paid-only. Second, Hydrogen’s price is flat across the full 1-million-token context window: there is no surcharge for long prompts. That flat pricing is a deliberate choice for agent workloads, where a single task can drag in a whole repository before it writes a line of code.
If you already read my Gemini 4 Argon guide, Hydrogen is the complement to that model, not its replacement. Argon is the one you reach for when a problem needs sustained reasoning; Hydrogen is the one you reach for when the plan is clear and the work needs doing quickly and cheaply.
What actually changes for everyday users?
Most people will never pick a model by name. For them, Hydrogen arrives silently, inside products they already use:
- Google Search: Hydrogen is now the default model behind AI Mode and is used for the hardest questions in AI Overviews, according to Google’s Search team. Google calls this its biggest model upgrade to Search yet. Expect answers that handle multi-part questions, typos, and very specific requests better, with more visual, scannable responses and a wider fan-out to relevant links.
- Gemini app: Hydrogen is available in the model dropdown for Google AI Plus, Pro, and Ultra subscribers. Free users get it through automatic model selection, when Google judges it the right model for the prompt.
- Google AI Studio and the Gemini API: available on all tiers, including the free tier, under the model ID
gemini-4-hydrogen. - Google Sheets: Hydrogen is coming to Gemini in Sheets, starting with Google AI Pro and Ultra subscribers.
- Google Antigravity: available to paid Pro and Ultra subscribers, alongside Orchestrator.
One practical catch in the Gemini app: your context window depends on your plan. Free accounts get 32,000 tokens of context, Plus gets 128,000, and only Pro and Ultra get the full 1 million. So the headline 1M window is real on the API, but in the app it is a paid feature — worth knowing before you paste a 200-page document into a free account and wonder why it gets truncated.
The Discover angle here is simple: this is the model now answering the hardest questions in Search. If your queries have felt noticeably better at multi-part questions this week, Hydrogen is a plausible reason why.
How much does Gemini 4 Hydrogen cost?
Hydrogen’s pricing is the easiest part of this launch to verify, because Google publishes it twice: once in the announcement and once on the Gemini API pricing page, which I checked before writing. The rates below are the Standard paid tier, in US dollars per million tokens. Thinking tokens are billed as output, which matters more than it sounds — more on that in the thinking-levels section.
| Tier | Input (per 1M) | Output, incl. thinking (per 1M) | Cached input (per 1M) |
|---|---|---|---|
| Free tier | No charge | No charge | No charge |
| Standard (paid) | $1.50 | $7.50 | $0.075 |
| Batch / Flex | $0.75 | $3.75 | $0.0375 |
| Priority | $2.70 | $13.50 | $0.135 |
Worked example. Say an agent task sends 100,000 input tokens (a chunk of a codebase plus instructions) and produces 10,000 output tokens including thinking. On the Standard tier that costs 100,000 × $1.50/1,000,000 = $0.15 of input plus 10,000 × $7.50/1,000,000 = $0.075 of output — about $0.23 for the task. Run the same task 100 times and the bill is about $23. The same task on Argon at its post-November-30 price ($4 in / $20 out) would cost about $0.60 — roughly 2.7 times more. That is the arithmetic behind Google’s “use Hydrogen for execution, Argon for reasoning” pitch, and it is why the model matters to anyone running agents at volume.
Three price changes around Hydrogen are worth putting on your calendar. First, Carbon and Argon’s introductory $2/$10 rate ends November 30, 2026 — from December 1 they cost $4/$20, which makes Hydrogen less than half their price. Second, Gemini 3.8 Flash, the model Hydrogen replaces, currently costs $0.75/$3.75 through December 31, 2026, then rises to Hydrogen’s $1.50/$7.50 from January 1, 2027. In other words, switching from Flash to Hydrogen doubles your per-token rate today, and costs the same from January. Third, the free tier has a condition: content sent on the free tier may be used to improve Google’s products. On the paid tier it is not. That is a real tradeoff if you are sending proprietary code through the free tier.
One more cost warning from independent testing: because thinking tokens are billed as output, and Hydrogen thinks at high unless you tell it otherwise, output can make up 89% to 99.7% of a call’s cost in AI Studio tests. If you are cost-sensitive, set the thinking level explicitly on every request — low or medium for short, high-volume calls, high for coding and agents, xhigh only where high falls short.
How fast and how smart is it, really?
Every number in this section comes from Google’s announcement and model card, and I am labeling it that way on purpose: these are vendor-reported results on benchmarks Google chose. Independent verification takes weeks, and I would rather show you Google’s claims with the caveats attached than dress them up as settled fact.
| Benchmark | Hydrogen result | What it means |
|---|---|---|
| Artificial Analysis Intelligence Index | 51 at default high thinking; 54 at xhigh | Edges out Argon (50) on this composite index — Hydrogen’s headline result. |
| Terminal-Bench 4.1 | 57% vs 18.2% for Gemini 3.8 Flash | More than triples Flash on agentic terminal work; fabricated claims cut by more than 5x. |
| Agent Arena | Ranks #5 | Live benchmark of open-ended, multi-step agent tasks. |
| HLE-Diamond | 58.8%, leads its class | Hard, expert-level reasoning questions. |
| Vals Tax Agent Bench | 52.3%, ranks #1 overall | Domain-specific tax agent work. |
| Vals Legal Research Bench | 55.8%, ties for #1 | Legal research agent work. |
| Multimodal Knowledge Agent (MKA) | 53.4% | End-to-end knowledge work over complex visual files. |
| Agents’ Last Exam v1.1 | 71.6% | End-to-end professional workflows. |
Google also claims Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking level, and that partners saw it run 2 to 2.8 times faster than Flash on complex graphs, SQL, and data-heavy enterprise workflows while using 12% to 23% fewer tokens.
Now the honest counterweight. The model card’s cheating-adjusted Terminal-Bench 4.1 table puts Claude Sonnet 5.5 at 63.9%, above Hydrogen’s 57.0% — so Hydrogen is not the best agent model on that benchmark once you adjust for cheating; it is the best in its price class. And the Artificial Analysis lead over Argon is one point, 51 versus 50, on a composite index — a real lead, but a narrow one, and it comes at the default thinking level where Hydrogen is tuned to shine. The right reading is not “Hydrogen beats everything.” It is “Hydrogen is remarkably close to the frontier models on agent work, at roughly a third of their post-November price.” If you want the coding comparison in more depth, my Claude vs Gemini for coding post is the natural companion to this one.
On safety, Google reports Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79% versus 3.8 Flash, cuts misaligned actions and task shortcutting by 12 times, is up to 2 times less likely to bypass permission boundaries on blocked tasks, and is more than 2.3 times more resilient to indirect prompt injection — hidden instructions planted in content an agent reads — on the Gray Swan benchmark. The model card adds a wrinkle worth knowing: unjustified refusals on benign prompts rose 5.57 percentage points versus Flash, so Hydrogen may say no more often to requests it could safely fulfill. Google also says Hydrogen reaches the alert thresholds (but not the critical capability levels) for biological-information and cybersecurity risk, with fuller safety documentation still to come.
What are the four thinking levels?
Hydrogen is the first Gemini model with four configurable thinking levels, and they are the main control you have over the speed-versus-quality tradeoff. Thinking level sets how much reasoning the model does before answering — and because thinking tokens are billed as output, it also sets much of your bill.
| Thinking level | Best for | Cost and speed note |
|---|---|---|
| low | Short, simple, high-volume calls — classifications, quick answers, bulk processing | Cheapest and fastest; least reasoning. |
| medium | Everyday writing, summarization, straightforward Q&A | A balanced default for non-coding work. |
| high (default) | Coding, debugging, tool-using agents, multi-step tasks | The default. This is where the 51-point AA Index score was measured. |
| xhigh (new) | The hardest problems where high falls short | Reaches 54 on the AA Index; the most expensive per call. |
The practical rule: leave it on high for agent and coding work, drop to low or medium for anything repetitive, and save xhigh for the one task in ten that actually needs it. Google’s own guidance is the same — you only wait and pay for the reasoning depth a task requires. In the API, the setting is thinking_level; in Antigravity’s CLI it is --effort xhigh, which the app labels “Extra.”
How to try Gemini 4 Hydrogen
You probably do not need to do anything to meet Hydrogen — if you use Google Search, it is already answering your hardest questions. But if you want to try it deliberately, here are the routes, in order of ease:
- Google Search (no setup): just search. Hydrogen is the default model behind AI Mode and the hardest AI Overview questions for all users, globally.
- Gemini app: open gemini.google.com and pick Gemini 4 Hydrogen from the model dropdown (Plus, Pro, and Ultra subscribers). On a free account, it is used through automatic model selection when Google judges it the right fit for your prompt.
- Google AI Studio (free tier, no billing): go to AI Studio, start a new chat, and select Gemini 4 Hydrogen. This is the best place to test the thinking levels, because you can switch between low, medium, high, and xhigh on the same prompt and watch the difference in depth and latency.
- Gemini API: use the model ID
gemini-4-hydrogenin your existing Gemini API code. If you are migrating fromgemini-3.8-flash, Google’s own guidance is to swap the model ID, raise any lowmax_output_tokenscap to 65,536, and remove explicit-caching calls — Hydrogen supports implicit caching only. - Google Antigravity (paid Pro/Ultra): Hydrogen is available in the model picker, and Orchestrator — the feature that hands planning to Carbon and execution to Hydrogen in one session — is available to Ultra subscribers, and to Pro subscribers in an Argon version.
A fair test to run yourself: take a prompt you already know the answer quality of — a code fix, a research summary, a document rewrite — and run it in AI Studio at low, then high, then xhigh. The jump from low to high is usually obvious; the jump from high to xhigh is where you learn whether your task actually needs the deepest level. If you do run that test, tell the model to show its work on a task with a verifiable answer, and check the answer yourself — that is the only benchmark that matters for your work. For research workflows specifically, my Gemini Deep Research guide and my Perplexity research guide both still apply: Hydrogen changes the engine, not the method.
What are the limits and honest caveats?
- Knowledge cutoff is March 2026. Hydrogen’s training knowledge ends in March 2026, and the model card notes that in some domains it may be less current. Anything after that — new prices, new laws, new product launches — needs live search or grounding, not the model’s memory.
- All headline benchmarks are Google’s. The 2.5x speed claim, the Terminal-Bench tripling, and the 51-point index score come from Google’s own evaluations. Third-party benchmarks will either confirm them or not over the coming weeks; treat them as a strong signal, not a verdict.
- It over-refuses more than Flash. Unjustified refusals on benign or borderline prompts rose 5.57 percentage points in Google’s own safety evaluation. If Hydrogen declines something reasonable, rephrase or drop a thinking level before assuming the request is actually disallowed.
- The free tier is not free of conditions. Content sent on the Gemini API free tier may be used to improve Google’s products, and the free tier is rate-limited (Google publishes no per-minute limits for this model). Do not send proprietary code or sensitive documents through it.
- It can still hallucinate. The model card lists hallucination as a known limitation, and notes occasional slowness or timeouts and higher token use at higher thinking levels. Hydrogen cuts fabricated claims by more than 5x versus Flash — which is a big improvement on a bad baseline, not a guarantee.
- Orchestrator is not for everyone yet. The multi-model routing that makes Carbon-plus-Hydrogen sing lives in Antigravity, which needs a paid Pro or Ultra subscription. API users can build the same pattern by hand, but Google does not do it for you outside Antigravity.
None of this makes Hydrogen a bad model. It makes it a real one — fast, cheap, and genuinely good at the work it was built for, with the same rough edges every fast model ships with. The models that look perfect on launch day are usually the ones nobody has stress-tested yet.
Should you switch from Gemini 3.8 Flash?
If you are on 3.8 Flash today, the answer is yes, with one condition: test before you migrate anything that matters. Google itself calls Hydrogen the “recommended upgrade from Gemini 3.8 Flash and earlier Gemini 3 and 2.5 models,” and the numbers support that — 2.5x faster, triple the Terminal-Bench score, a fifth of the fabricated claims, and a 1-million-token window with no long-prompt surcharge. The only cost is that per-token rates double until January 1, 2027, when Flash’s own price rises to match.
The switch is also cheap to test: one model-ID change in the API, or one dropdown selection in AI Studio. Run your ten most common prompts on both models for a week, compare the outputs and the bill, and keep whichever wins on your actual work. For most people that will be Hydrogen — but “most people” is not a measurement, and your ten prompts are.
If you are choosing between Hydrogen and Argon instead, the rule is simpler: Argon plans, Hydrogen executes. Use Argon (or Carbon) for the hard reasoning — the architecture, the tricky bug, the ambiguous brief — and let Hydrogen do the implementation, the tests, and the verification passes. That split is what Orchestrator automates, and it is the workflow Google is betting the whole Gemini 4 family on. My ChatGPT vs Gemini for research comparison is the place to go if you are weighing Gemini against the other labs rather than against itself.
Gemini 4 Hydrogen is Google’s fast, lower-cost Gemini 4 model, launched on November 12, 2026. It is built for everyday coding and multi-step agentic workflows — writing code, running tests, verifying changes, and carrying tasks to completion — and it replaces Gemini 3.8 Flash as Google’s default workhorse model. It is available in the Gemini app, Google Search, Google AI Studio, and the Gemini API.
Yes, in three places: in Google Search for all users, in the Gemini app through automatic model selection on a free account, and on the free tier of the Gemini API and Google AI Studio. The free tier is rate-limited, and content sent on it may be used to improve Google’s products. On the paid API tier, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens.
Carbon and Argon are Google’s frontier models for deep, long-horizon reasoning on hard problems. Hydrogen is the fast execution model — it does the coding, testing, and multi-step work quickly and cheaply. Hydrogen costs $1.50 per million input tokens versus $2 (rising to $4 on December 1) for Argon and Carbon, and it is the only one of the three with a free tier. Google’s Orchestrator feature pairs them: Carbon plans, Hydrogen executes.
Hydrogen offers four thinking levels: low, medium, high (the default), and a new xhigh tier. Low is for short, high-volume calls; medium for everyday writing and summarization; high for coding and agent work; xhigh for the hardest problems where high falls short. Thinking tokens are billed as output, so a higher level costs more per call — set the level explicitly instead of leaving every request on high.
On the Standard paid tier, $1.50 per million input tokens and $7.50 per million output tokens, with cached input at $0.075 per million. Batch and Flex cost half that ($0.75 / $3.75), and Priority costs $2.70 / $13.50. The price is flat across the full 1-million-token context window, with no surcharge for long prompts. A task using 100,000 input tokens and 10,000 output tokens costs about $0.23.
Yes, for most uses — Google calls Hydrogen the recommended upgrade from 3.8 Flash, and it is 2.5 times faster with triple the Terminal-Bench score and far fewer fabricated claims. The only tradeoff is price: Hydrogen costs twice as much per token as Flash until January 1, 2027, when Flash’s price rises to match. Test your ten most common prompts on both models before migrating anything important.