Gemini 4 Hydrogen pricing — free, but you cannot pick it; the API costs twice 3.8 Flash
Gemini 4 Hydrogen, released by Google on November 12, 2026 (US time), costs $1.50 per million input tokens and $7.50 per million output tokens on the API. The same rate applies across the full 1-million-token context. Cached input is $0.075 and the batch and flex tiers are half price at $0.75 and $3.75. That is twice the current introductory price of Gemini 3.8 Flash ($0.75 and $3.75), which rises to the same $1.50 and $7.50 on January 1, 2027. Free users of the Gemini app get Hydrogen only when automatic model selection picks it and cannot choose it themselves. In Google Search it applies to all users. Developers can use it on the free tier of Google AI Studio and the Gemini API, where prompts may be used to improve Google products. On an independent index it scores 51 at its default setting, one point above Argon
The three lines
- Price — $1.50 input and $7.50 output per million tokens, $0.075 cached, batch at half. Twice the current 3.8 Flash rate
- Free access — Gemini app by automatic selection only, Search for everyone. The free API tier may use prompts for product improvement
- Performance — independent index 51 (default) and 54 (top). At the top setting a task costs $3.16, 3.3 times GPT-6.1 Terra's $0.96
Key questions
- Gemini 4 Hydrogen pricing
- **$1.50 per million input tokens and $7.50 per million output tokens. There is no surcharge for long prompts and no end date attached to the rate.** | Tier (per 1M tokens) | Input | Output (includes thinking tokens) | Cached input | |---|---|---|---| | Standard | $1.50 | $7.50 | $0.075 | | Batch | $0.75 | $3.75 | $0.0375 | | Flex | $0.75 | $3.75 | $0.0375 | | Priority | $2.70 | $13.50 | $0.135 | | Free tier | Free of charge | Free of charge | — | From the Google AI Studio price list, viewed November 16, 2026.
- Is Gemini 4 Hydrogen free
- **There are three free routes. Free users of the Gemini app cannot select Hydrogen directly; they get it only when automatic model selection chooses it.** | Where | Free users | Condition | |---|---|---| | Gemini app | Only when automatic selection picks it | Most prompts are answered by Flash-Lite | | Google Search (AI Mode and more) | All users | Rolling out, usage limits not published | | Google AI Studio and Gemini API | Free tier available | Prompts may be used to improve products | | Antigravity | Not included | Requires Google AI Pro or Ultra | | Direct selection | Not possible | Starts with Google AI Plus ($4.99 a month in the US) |
- Gemini 4 Hydrogen vs Argon vs Carbon
- **Hydrogen is the cheapest and fastest model in the family, built for execution. Carbon and Argon are for long-running reasoning and have no free API tier.** | Item | Carbon | Argon | Hydrogen | |---|---|---|---| | Input / output per 1M tokens | $2 / $10, then $4 / $20 from December 1 | $2 / $10, then $4 / $20 from December 1 | $1.50 / $7.50 | | Free API tier | No | No | Yes | | Gemini app | Ultra subscribers | Pro and Ultra subscribers | Everyone (free users via automatic selection) | | Artificial Analysis index | 60 (high) | 50 (high) | 51 (high), 54 (xhigh) | | Time per index task | 33.8 min | 30.9 min | 8.4 min (high), 14.8 min (xhigh) |
Google opened the third model in its Gemini 4 family to everyone on day one. On Thursday, November 12, 2026 (US time), Google released Gemini 4 Hydrogen. It follows Argon in October and Carbon, which reached paying customers two days earlier on November 10. Those two are sold only to paying users and have no free API tier. Hydrogen launched with free Gemini app users, Google Search users and free-tier developers in Google AI Studio all included.
The API price is $1.50 per million input tokens and $7.50 per million output tokens. Google's headline claim is that this cheaper model outscores Argon, the model one step above it, on an independent index. Set the numbers side by side and three qualifications appear. The price is twice what Gemini 3.8 Flash costs now. Free users cannot select Hydrogen themselves. And against a competitor whose list price is a third lower, Hydrogen's cost per finished task is 3.3 times higher. This article goes through the price list, Hydrogen's place in the family, competing models, worked cost examples, the free routes and their limits, availability in South Korea as a test case for non-US markets, who measured which benchmark, and what Gemini 3.8 Flash users should do.
1. The price list — $1.50 and $7.50 per million tokens
These are the rates on the Google AI Studio price list. The model ID is gemini-4-hydrogen.
| Tier (per 1M tokens) | Input | Output (includes thinking tokens) | Cached input | Free tier |
|---|---|---|---|---|
| Standard | $1.50 | $7.50 | $0.075 | Yes |
| Batch | $0.75 | $3.75 | $0.0375 | No |
| Flex | $0.75 | $3.75 | $0.0375 | No |
| Priority | $2.70 | $13.50 | $0.135 | Yes |
| Contributor (invite-only private preview) | $0.75 | $3.75 | $0.0375 | Yes |
Five points matter when reading it.
First, there are no prompt-length tiers. Google's announcement says the same rate applies across the full 1-million-token context window. Claude Haiku 5.5 charges five times more once a prompt passes 100,000 tokens, and OpenAI applies higher rates above 272,000 input tokens. For jobs that load long documents or a whole codebase, that difference can outweigh the list price.
Second, the rate carries no end date. The price list attaches November 30 to the Argon and Carbon rates and December 31 to the Gemini 3.8 Flash rates. The Hydrogen rows have no date.
Third, output charges include thinking tokens, the tokens a model spends reasoning before it writes an answer. Hydrogen has four thinking levels: low, medium, high (the default) and xhigh. Higher levels produce more output tokens. A visible answer of 500 tokens is billed as more than 500.
Fourth, caching is implicit only. The price list says explicit caching, where a developer creates and pins a cache, is not supported. Cached input costs $0.075, one twentieth of the list input price. According to Data Studios, implicit caching applies to prompts of at least 4,096 tokens; we found that condition in one source only.
Fifth, the Contributor tier halves the price in exchange for your data. The price list says content on this tier is used to improve products whether or not you pay. On the Standard tier, paid usage is not. Meta introduced the same kind of pricing in September.
Batch and flex are for work that does not need an immediate response; priority is for latency-sensitive work. Grounding answers in Google Search is billed separately: the price list gives 5,000 free requests a month across Gemini 3 and later models, then $14 per 1,000 requests.
2. Where Hydrogen sits in the Gemini 4 family
Google's announcement, published under the name of Koray Kavukcuoglu, senior vice president at Google DeepMind, describes Hydrogen as the model for fast everyday execution and positions Carbon and Argon for deep, long-horizon reasoning. SiliconANGLE counted three Gemini 4 models shipped in just over six weeks.
| Item | Carbon | Argon | Hydrogen |
|---|---|---|---|
| Broad release (US time) | November 10 (defenders only from October 28) | October 13 | November 12 |
| Input / output per 1M tokens | $2 / $10, then $4 / $20 from December 1 | $2 / $10, then $4 / $20 from December 1 | $1.50 / $7.50 |
| Cached input | $0.10, then $0.20 from December 1 | $0.10, then $0.20 from December 1 | $0.075 |
| Free API tier | No | No | Yes |
| Gemini app | Ultra subscribers | Pro and Ultra subscribers | Everyone; free users through automatic selection only |
| Thinking levels (per Chrome Geek) | Default medium, top high | Default medium, top high | Default high, top xhigh |
| Artificial Analysis index | 60 (high) | 50 (high) | 51 (high), 54 (xhigh) |
| Time per index task | 33.8 min | 30.9 min | 8.4 min (high), 14.8 min (xhigh) |
| Intended work (Google's description) | The hardest long-running tasks | Most software and knowledge work | Everyday coding, tool calls, multimodal tasks |
Start with price. Through November 30 Hydrogen is 25% cheaper than Argon and Carbon. When those two move to $4 and $20 on December 1, Hydrogen will cost 37.5% of their price. The Next Web's statement that Argon costs more than twice as much refers to the rates from December 1.
On scores, Google's claim and the independent figure agree. The Artificial Analysis page shows Hydrogen at 51 on its default high setting and Argon at 50 on high, a one-point gap. At xhigh Hydrogen reaches 54. On the same index Carbon scores 60. Hydrogen outscores the model one step above it, not the top of the family.
The speed gap is clearer. On the same page a single index task took 8.4 minutes for Hydrogen on high and 30.9 minutes for Argon. Output speed came in at around 170 tokens per second at every setting.
Google also shipped a way to combine the two kinds of model. Orchestrator, in its Antigravity coding tool, switches between Carbon and Hydrogen within one job, with Carbon planning and coordinating while Hydrogen runs worker and verifier agents. Google says that on Terminal-Bench 4.0 the combination came close to Carbon's success rate while using 24% less quota and finishing 19% faster on average. That is Google's own measurement. According to Chrome Geek, Orchestrator opened first to Ultra subscribers and Gemini for business only.
3. Against competitors — list price and cost per task tell different stories
These are list prices for models in Hydrogen's range and for those above and below it. The Google, OpenAI and Anthropic figures were rechecked on each company's price list on November 16, 2026.
| Model | Input | Output | Cached input | Long-prompt surcharge |
|---|---|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 | $0.20 | — |
| Gemini 4 Carbon and Argon | $2.00, then $4.00 | $10.00, then $20.00 | $0.10, then $0.20 | None |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 | — |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 | Higher above 272,000 input tokens |
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 | None |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 | Higher above 272,000 input tokens |
| Gemini 3.8 Flash | $0.75, then $1.50 from January 1, 2027 | $3.75, then $7.50 | $0.075, then $0.15 | None |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 | $0.50 / $2.50 above 100,000 tokens |
| Meta Muse Ember | $0.10 | $0.50 | $0.004 | $0.30 / $1.50 above 200,000 tokens |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | — |
On list price Hydrogen sits in the middle. It is 25% cheaper than Sonnet 5.5 and GPT-6.1 Sol and 50% more expensive than GPT-6.1 Terra. The gap to the $0.10 cluster of Haiku 5.5, Muse Ember and Luna is fifteenfold. Hydrogen is not the cheapest model. Google's cheapest line is Flash-Lite; Hydrogen is the mid-priced tier above it.
What a customer pays is list price multiplied by tokens used. Artificial Analysis separately measures the cost of completing one task from its index. This is its leaderboard as viewed on November 16, 2026.
| Model (setting) | Index score | Cost per task | Time per task |
|---|---|---|---|
| Claude Fable 5.2 (max) | 65 | $26.54 | 44.5 min |
| Gemini 4 Carbon (high) | 60 | $5.48 | 33.8 min |
| Claude Sonnet 5.5 (max) | 55 | $10.02 | 33.3 min |
| Gemini 4 Hydrogen (xhigh) | 54 | $3.16 | 14.8 min |
| GPT-6.1 Terra (max) | 51 | $0.96 | 14.1 min |
| Gemini 4 Argon (high) | 50 | $4.97 | 30.9 min |
| Claude Haiku 5.5 (max) | 36 | $0.41 | 14.8 min |
| Meta Muse Ember (max) | 36 | $0.14 | 3.3 min |
| GPT-6 Luna (max) | 28 | $0.15 | 12.2 min |
The table reads in two directions.
Looking up, Hydrogen is cheap. Sonnet 5.5 at its top setting scores 55 at $10.02; Hydrogen at its top setting scores 54 at $3.16. One point apart, at less than a third of the cost and less than half the time. That is a far wider gap than the 25% difference in list price. Against Argon, Hydrogen scores four points higher at 36% lower cost.
Looking down, Hydrogen is expensive. GPT-6.1 Terra at its top setting scores 51 at $0.96. Hydrogen's list price is 1.5 times Terra's, but its cost per task is 3.3 times higher, for three more points. Hydrogen uses more tokens on the same task. Terra's 51 equals Hydrogen's score at its default setting, but the leaderboard shows no cost per task for Hydrogen at that setting, so a like-for-like comparison at the same score is not possible.
The Carbon and Argon costs in this table differ from the $7.76 and $9.95 recorded in our November 12 article. They are readings of the same leaderboard on different days and we could not confirm why they changed. Cost per task is not a fixed number; it moves with the measurement date and the index version.
4. Three worked cost examples
These are arithmetic: list price times assumed tokens. Models use different numbers of tokens for the same job, so real invoices will differ. Use them to gauge scale.
Example 1. A customer support chatbot handling 100,000 conversations a month. Assume 2,000 input tokens per request (instructions plus the question) and 500 output tokens. That is 200 million input tokens and 50 million output tokens a month.
| Model | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Gemini 4 Hydrogen | $300 | $375 | $675 |
| Gemini 3.8 Flash (through December 31) | $150 | $187.50 | $337.50 |
| Gemini 3.8 Flash (from January 1, 2027) | $300 | $375 | $675 |
| Gemini 4 Argon (through November 30) | $400 | $500 | $900 |
| Gemini 4 Argon (from December 1) | $800 | $1,000 | $1,800 |
| GPT-6.1 Terra | $200 | $250 | $450 |
| Claude Sonnet 5.5 | $400 | $500 | $900 |
| Claude Haiku 5.5 | $20 | $25 | $45 |
What the table leaves out is thinking tokens. Hydrogen defaults to high. If it spends 1,500 thinking tokens before writing a 500-token answer, billed output becomes 2,000 tokens per request; the output charge rises from $375 to $1,500 and the total from $675 to $1,800. The 1,500 figure is an assumption for illustration. For simple support replies, lowering the thinking level to low or medium is the largest lever on cost. The trade-off is visible on the Artificial Analysis index, where Hydrogen scores 27 on low, 41 on medium and 51 on high.
Example 2. One coding agent session of 40 calls. Assume each call sends 60,000 input tokens, of which 80% (48,000 tokens) repeats the previous call and is read from cache. Each call returns 3,000 output tokens.
| Item | Tokens | Rate per 1M | Cost |
|---|---|---|---|
| Cached input | 1.92 million | $0.075 | $0.144 |
| New input | 480,000 | $1.50 | $0.72 |
| Output | 120,000 | $7.50 | $0.90 |
| Total | $1.764 | ||
| If nothing hits the cache | 2.4 million input, 120,000 output | $4.50 |
With caching the same session falls from $4.50 to about $1.76, a 61% reduction. Most of an agent's input is repeated context, which makes the cached rate as important as the list price. Hydrogen's cached rate is 5% of list input. The caveat is that implicit caching does not guarantee a hit. Developers cannot create and pin a cache, so the hit rate depends on keeping the start of the prompt identical from call to call.
Example 3. Summarizing 10,000 documents overnight, by tier. Assume 8,000 input tokens per document and an 800-token summary. That totals 80 million input tokens and 8 million output tokens.
| Tier | Input cost | Output cost | Total |
|---|---|---|---|
| Standard | $120 | $60 | $180 |
| Batch or flex | $60 | $30 | $90 |
| Priority | $216 | $108 | $324 |
The same job costs between $90 and $324 depending on the tier, a 3.6-fold spread. Work that does not need an immediate answer halves its cost simply by going through batch.
5. How to use it for free, and the limits
The word free covers three different routes with different conditions.
| Route | What is free | What is not, and conditions | Source |
|---|---|---|---|
| Gemini app | You get Hydrogen's answer when automatic model selection picks it | No direct selection. Most prompts go to Flash-Lite | Google blog, Transclew |
| Google Search | Applies to all users in AI Mode and elsewhere | Rolling out. Usage limits not published | Google Search blog |
| Google AI Studio | Pick the model in a browser and test it | Free-tier prompts may be used to improve products | Google price list |
| Gemini API free tier | Free calls on the standard tier | Request limits apply (numbers not published). No free batch or flex | Google price list, rate-limits page |
| Antigravity, Google Sheets | Nothing | Google AI Pro, Ultra or a business subscription | Google blog |
The Gemini app. Google's announcement says free users get Hydrogen through automatic model selection when it fits the prompt. According to Transclew and GSM Go Tech, free users cannot choose Hydrogen from the model menu; with automatic selection on, most prompts go to Flash-Lite and only those needing more reasoning may be passed to Hydrogen. With automatic selection off, every reply comes from Flash-Lite. Google has not said how often free users are routed to Hydrogen. Both outlets say the model that answered can be checked from the More menu under a reply.
Selecting Hydrogen directly requires a paid plan. US prices below are as reported by GSM Go Tech and The AI Career Lab; the Korean prices are from Google's Korean subscriptions page and are included as an example of a non-US market.
| Plan | US price per month | South Korea price per month | Hydrogen | Usage allowance (Google's wording) | Context window in the app |
|---|---|---|---|---|---|
| Free | $0 | 0 won | Automatic selection only | Baseline | 32,000 tokens |
| Google AI Plus | $4.99 | 7,500 won | Selectable | 2x free | 128,000 tokens |
| Google AI Pro | $19.99 | Not shown on the page | Selectable, plus Argon | 4x free | 1 million tokens |
| Google AI Ultra | $99.99 or $199.99 (per VentureBeat, cited in our November 12 article) | 119,000 or 300,000 won | Selectable, plus Argon and Carbon | Higher than Pro | 1 million tokens |
The subscriptions page says allowances refresh every five hours until a weekly limit is reached. The 2x and 4x figures are ratios, not prompt counts, and Google's Help Center notes that limits can change without notice. The context window figures come from that Help Center article, which did not yet mention Hydrogen. Transclew reports that the app offers low, medium and high thinking levels but not xhigh, and that a paying user who selects Hydrogen without choosing a level gets low. Given that the index score is 27 on low and 51 on high, that setting is worth checking.
The developer free tier. Google's price list marks Hydrogen's standard tier as free of charge on the free tier, something Argon and Carbon do not have. The price is data. The list says content is used to improve products on the free tier and not on the paid tier, which is the reason not to put company code or customer information through the free tier. Google's documentation gives no per-minute or per-day request numbers; the rate-limits page says limits vary by model and points to the AI Studio console.
6. Availability outside the US — South Korea as a test case
Google describes the rollout as global and has published no country schedule. We checked what can be confirmed for one non-English market, South Korea, where this publication is based.
| Item | Status | Evidence |
|---|---|---|
| Gemini app (Korean accounts) | Hydrogen appears in the plan table | Google's Korean subscriptions page lists the model under the free plan and AI Plus (7,500 won a month) |
| AI Mode in Google Search (Korean) | Korean is a supported language | AI Mode added Korean on September 8, 2025. Google's Search post says answers improve across all supported AI Mode languages |
| Gemini API and AI Studio | South Korea is an available region | It appears on Google's region list |
| Billing currency | The API is priced in US dollars | The price list shows dollars only |
| Actual rollout date | Not confirmed | Google gives no per-country dates |
The model itself handles languages other than English as input and output. The model card says it accepts text, images, audio and video, has a context of up to 1 million tokens shared between input and output, returns up to 64,000 output tokens by default in a synchronous call and longer output through a background mode, and has a knowledge cutoff of March 2026. We found no published figure for Korean-language quality specifically. The only language-related number on the model card is a multilingual safety evaluation that improved by 1.10 percentage points over Gemini 3.8 Flash.
Token-based pricing also lands differently by language. The same sentence often produces more tokens in Korean, Japanese or Chinese than in English, so the same list price yields a higher effective cost per sentence.
7. Benchmarks — sorted by who measured them
The numbers that came with the Hydrogen launch have three kinds of source: values that independent organizations publish on their own pages, a comparison table in Google's model card, and Google's own measurements quoted in its announcement.
| Metric | Hydrogen | Comparison | Measured by |
|---|---|---|---|
| Artificial Analysis Intelligence Index | Low 27, medium 41, high 51, xhigh 54 | Argon 50, Carbon 60, Sonnet 5.5 55, GPT-6.1 Terra 51 | Artificial Analysis (independent) |
| Vals Index | 68.49% (±0.99), 6th of 37 | Fable 5.2 71.31% | Vals AI (independent), at xhigh, $9.84 per test |
| Vals Tax Agent Bench | 52.30% (±3.60), 1st of 85 | Carbon 51.78% | Vals AI (independent) |
| Vals Legal Research Bench | 53.85%, 6th of 94 on the Vals page | Google announced 55.8%, tied for first | The two differ |
| Terminal-Bench 4.0 (Vals run) | 54.04%, 11th of 64 | — | Vals AI (independent) |
| Terminal-Bench 4.1 | 57.0% | 3.8 Flash 18.2%, Sonnet 5.5 63.9% | Google model card |
| HLE-Diamond (reasoning) | 58.8% | Sonnet 5.5 53.3%, Terra 47.3% | Google model card |
| AutomationBench-AA (work automation) | 76.1% | Sonnet 5.5 71.8%, Terra 60.2% | Google model card |
| Agents' Last Exam v1.1 | 71.6% | Terra 69.8%, Sonnet 5.5 60.4% | Google model card |
| OSWorld 2.1 (computer use, offline subset) | 77.6% | Sonnet 5.5 83.9%, Terra 74.2% | Google model card |
| Terminal-Bench-Science 0.1 | 51.4% | Sonnet 5.5 59.9%, Terra 58.3% | Google model card |
| Task completion speed | 2.5x Gemini 3.8 Flash | 2x to 2.8x in partner tests | Google announcement |
The two independent measurements point the same way. Hydrogen scores high for its price and below the top models. Vals AI places it sixth overall, and according to The Next Web describes it as the cheapest and fastest model in its top six. Vals also characterizes its coding results as mid-pack; it ranks 11th of 64 on Terminal-Bench 4.0.
The model card's comparison table sets Hydrogen, Sonnet 5.5 and GPT-6.1 Terra against ten benchmarks. Hydrogen is highest in seven rows and another model is higher in three. The losing rows are the notable ones. On Terminal-Bench 4.1, the benchmark Google puts at the front of its announcement, Hydrogen rose from Gemini 3.8 Flash's 18.2% to 57%, yet Sonnet 5.5 in the same table scores 63.9%. Sonnet 5.5 also leads on computer use and on the science terminal benchmark. A threefold jump over the previous model and a loss to a competitor are both true on the same test.
One row shows Google's announcement and an independent page disagreeing. Google said Hydrogen tied for first on the Vals Legal Research benchmark at 55.8%; the Vals page opened on November 16, 2026 shows 53.85% and sixth place. Leaderboards move as new models are added and runs are repeated, so a figure quoted at launch needs rechecking days later.
On safety, the model card also includes an unfavorable number. Unjustified refusals, meaning refusals of harmless requests, rose by 5.57 percentage points over Gemini 3.8 Flash, particularly on multimodal prompts. Google's announcement says overreach in agent tasks fell by up to 81% and fabricated claims fell more than fivefold; those are internal evaluations, and the full safety report is listed as coming soon.
8. What Gemini 3.8 Flash users should do
Several articles say Hydrogen replaces Gemini 3.8 Flash. What Google's documents say is narrower. The announcement calls Hydrogen a major upgrade for teams migrating from 3.8 Flash and does not say 3.8 Flash is ending. Google's deprecations table lists gemini-3.8-flash with no shutdown date announced and no recommended replacement. There is no shutdown schedule. There is a price increase schedule.
| Date | What changes | Rate (input / output per 1M tokens) |
|---|---|---|
| November 12, 2026 | Hydrogen launches | $1.50 / $7.50 |
| November 30, 2026 | Last day of introductory pricing for Argon and Carbon | $2 / $10 |
| December 1, 2026 | Argon and Carbon move to standard pricing | $4 / $20 |
| December 31, 2026 | Last day of introductory pricing for 3.8 Flash | $0.75 / $3.75 |
| January 1, 2027 | 3.8 Flash moves to standard pricing | $1.50 / $7.50, cached input $0.15 |
| Not set | 3.8 Flash shutdown | No announcement |
On January 1, 2027 the list prices of 3.8 Flash and Hydrogen become identical, and 3.8 Flash's cached input rate of $0.15 will be twice Hydrogen's $0.075. After that date there is no price reason to stay on 3.8 Flash. Until December 31, 3.8 Flash costs half as much. The introductory rate and its end date were announced when 3.8 Flash launched in September.
For teams on 3.8 Flash, the to-do list looks like this.
| Action | Reason |
|---|---|
| Model the January 2027 invoice now | At the same usage, the 3.8 Flash bill doubles |
| Test Hydrogen on your own workload | Google says partners used 12% to 23% fewer tokens, but that is its own claim. Billed output, thinking tokens included, varies by job |
| Set the thinking level explicitly | The default is high. Simple work run at the default may generate unnecessary output tokens |
| Rework any explicit caching | Hydrogen supports implicit caching only, per Google's price list |
| Watch for more refusals | The model card reports unjustified refusals 5.57 points above 3.8 Flash |
| Send non-urgent work through batch | $0.75 input and $3.75 output, the same as 3.8 Flash's current standard rate |
The last row is worth calculating. Hydrogen's batch price equals 3.8 Flash's current standard price. For work that can wait for a response, a team can change models without paying more. Data Studios adds that on migration sampling parameters such as temperature are ignored and that the lowest thinking level, minimal, returns an error; we could not confirm those points in Google's documentation.
9. Frequently asked questions
Q. Is Gemini 4 Hydrogen free? In the Gemini app and Google Search it can be used without paying. Free app users get it only when automatic model selection picks it and cannot choose it. Developers can use the free tier of Google AI Studio and the Gemini API, where prompts may be used to improve Google's products.
Q. What does the API cost? $1.50 per million input tokens and $7.50 per million output tokens. Cached input is $0.075, batch and flex are half price, and priority is 1.8 times the standard rate. Output charges include thinking tokens.
Q. Is it cheaper than Gemini 3.8 Flash? No. Through December 31, 2026, 3.8 Flash costs half as much. From January 1, 2027 the two list prices are the same and Hydrogen's cached input is cheaper.
Q. Is it really better than Argon? On the Artificial Analysis index it leads by one point, 51 to 50, at each model's high setting. Google still recommends Argon and Carbon for long, deep reasoning. Carbon scores 60 on the same index.
Q. When does Gemini 3.8 Flash shut down? No date has been announced. Google's deprecations table lists none. The only confirmed change is the price increase on January 1, 2027.
Q. Is it available outside the United States? Google calls the rollout global. In South Korea, the market we checked, Hydrogen appears on the local subscriptions page, the country is on the API region list and AI Mode supports the language. Account-level rollout is gradual and no country schedule has been published.
Q. How can I tell which model answered? According to Transclew and GSM Go Tech, the Gemini app shows it under the More menu beneath a reply, in the response details or thinking steps. We did not test this ourselves.
10. What remains open and unconfirmed
| Open question | Current status |
|---|---|
| Request limits on the free API tier | No numbers in Google's documentation, only a pointer to the AI Studio console |
| How often free app users get Hydrogen | Not disclosed. Google says only that most prompts go to Flash-Lite |
| Prompt counts by plan | Only relative ratios of 2x and 4x. Google's own pages differ on the Ultra multiple |
| Gemini 3.8 Flash shutdown date | Not announced |
| Implicit caching threshold (4,096 tokens) | One report only |
| Whether batch pricing existed on launch day | It is on the price list now. One report says it was missing at launch |
| Cost per task at Hydrogen's default setting | Artificial Analysis shows only the xhigh figure of $3.16 |
| Movement in Carbon and Argon cost per task | Readings on November 12 and November 16 differ. Cause unconfirmed |
| Vals Legal Research ranking | Google says tied for first; the Vals page shows sixth |
| 2.5x speed, fivefold drop in fabricated claims | Google's own measurements. Method unpublished, no independent reproduction |
| Rollout dates by country, non-English quality | No schedule. No published language-specific quality figures found |
| Safety report | Listed as coming soon on the model card |
Hydrogen's price has to be read through three numbers: the list rate of $1.50 and $7.50, the number of output tokens the chosen thinking level generates, and the cache hit rate. The same model moves from 27 to 54 on the index depending on the setting, and its cost moves with it. The basis for choosing is not the price list but what your own workload costs at the level you run it.
Pricing and independent measurements for the top of the family are in our November 12 article on Gemini 4 Carbon pricing and availability. Competing releases from the same week are covered in the November 11 articles on Claude Fable 5.2 and Meta Muse Ember. The model whose price list Hydrogen inherits is covered in our September article on Gemini 3.8 Flash, and the pattern by which one year's top model becomes the next year's mid-priced tier is explained in What is a frontier model.
Sources
- Google Blog — Introducing Gemini 4 Hydrogen for Agentic Workflows (2026-11-12, updated 11-13)
- Google Blog (Search) — Gemini 4 Hydrogen Search upgrade (2026-11-12)
- Google AI Studio — Gemini API pricing (gemini-4-hydrogen, gemini-4-argon, gemini-4-carbon, gemini-3.8-flash; viewed 2026-11-16)
- Google AI Studio — Gemini API deprecations (viewed 2026-11-16)
- Google AI Studio — Gemini API rate limits (viewed 2026-11-16)
- Google AI Studio — Available regions for the Gemini API and AI Studio
- Google DeepMind — Gemini 4 Hydrogen model card
- Google — Gemini subscription plans (Korean page, viewed 2026-11-16)
- Google Help — Gemini app limits and upgrades for Google AI subscribers
- Google Blog (Search) — AI Mode expands to more languages (2025-09-08)
- Artificial Analysis — Gemini 4 Hydrogen model page (viewed 2026-11-16)
- Artificial Analysis — LLM Leaderboard (viewed 2026-11-16)
- Vals AI — Gemini 4 Hydrogen model page (viewed 2026-11-16)
- SiliconANGLE — Google releases Gemini 4 Hydrogen, a lower-cost model said to outscore Argon (2026-11-12)
- The Next Web — Google's Gemini 4 Hydrogen brings a fast, cheaper model to everyone (2026-11-12)
- Data Studios — Google Gemini 4 Hydrogen launch: API pricing, Gemini app, Search (2026-11-12)
- Chrome Geek — Gemini 4 Hydrogen (2026-11-12)
- Transclew — Gemini 4 Hydrogen is free, but Google picks when you get it (2026-11-15)
- iClarified — Google Launches Gemini 4 Hydrogen, Bringing Its Latest AI to Free Users (2026-11-12)
- GSM Go Tech — Gemini 4 Hydrogen Launches: Free Access, Google AI Plus Pricing and Key Features (2026-11)
- The AI Career Lab — What Is Gemini 4 Hydrogen, and Can Free Gemini Users Get It?
- OpenAI — API pricing (gpt-6.1-terra, gpt-6.1-sol, gpt-6-luna; viewed 2026-11-16)
- Anthropic — Claude API pricing (Fable 5.2, Sonnet 5.5, Haiku 5.5; viewed 2026-11-16)