Google launches Gemini 4 Hydrogen at $1.50/$7.50 with a 1M-token window

2 hours ago

10 min read

Google released Gemini 4 Hydrogen on November 12, 2026, adding a third model to the Gemini 4 family alongside Gemini 4 Carbon and Gemini 4 Argon. The company describes Hydrogen as its "daily driver for agentic workflows", a model tuned for fast execution of everyday coding, tool use and multimodal tasks, while Carbon and Argon remain the models for deep, long-horizon reasoning. The announcement was published on Google's blog at 18:00 UTC by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, and the rollout started the same day across developer, enterprise and consumer products.

The headline numbers for developers are a price of $1.50 per million input tokens and $7.50 per million output tokens, flat across the whole context window, with implicitly cached input billed at $0.075 per million tokens. The model, exposed in the Gemini API as `gemini-4-hydrogen`, works inside a single 1,048,576-token window shared by input, thinking and output, and it offers four thinking levels, with `high` as the default and a new `xhigh` tier at the top. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default setting, one point above Gemini 4 Argon, and 54 at `xhigh`.

Hydrogen also reaches ordinary users on launch day. It is rolling out globally in Google Search for all users, it appears in the model menu of the Gemini app, free Gemini users can receive it through automatic model selection, and paid Google AI Pro and Ultra subscribers get it in Google Antigravity. Google positions the model as the upgrade path for teams still running Gemini 3.8 Flash, and it claims tasks finish 2.5 times faster than on 3.8 Flash at the default setting.

··········

GEMINI 4 HYDROGEN AVAILABILITY IN THE GEMINI API, GEMINI APP AND GOOGLE SEARCH.

Google lists a broad set of surfaces for the first day. On the developer side, Hydrogen is available to all tiers of Google AI Studio and through the Gemini API, and the developer documentation confirms that a free tier exists, subject to rate limits that Google publishes on its separate rate-limit page. Android Studio users get the model on the paid tier. Businesses can select it in Gemini for business, either through the API or from the model dropdown, and enterprises also get it inside Antigravity.

On the consumer side, Hydrogen sits in the model dropdown of the Gemini app. Free users do not choose it directly by default; Google says they will reach it through automatic model selection when it fits the prompt, and that most free prompts will still be handled by a Flash-Lite model. Gemini in Google Sheets gets Hydrogen first for Google AI Pro and Ultra subscribers and eligible business subscribers. In Google Search, Hydrogen is rolling out globally for all users in every language that AI Mode supports, which makes Search the widest distribution channel of the launch.

........

Surface	Who gets Gemini 4 Hydrogen	Status on November 12, 2026
Gemini API and Google AI Studio	All tiers, free tier with rate limits	Available
Android Studio	Paid tier	Available
Gemini for business	Via API or model dropdown	Available
Google Antigravity	Google AI Pro and Ultra subscribers, enterprises	Available
Gemini app	Model dropdown; free users via auto selection	Rolling out
Gemini in Google Sheets	AI Pro, AI Ultra and eligible business plans	Rolling out
Google Search	All users, all AI Mode languages	Rolling out globally

........

Google's launch materials do not mention Google AI Plus, Vertex AI or the Gemini CLI, and they give no country-by-country schedule beyond the statement that the rollout is global. The company also does not state that Hydrogen replaces Gemini 3.8 Flash as a default anywhere; it describes Hydrogen as the model to migrate to.

··········

GEMINI 4 HYDROGEN API PRICING COMPARED WITH GEMINI 3.8 FLASH AND GEMINI 4 ARGON.

Hydrogen has a single price list with no prompt-length tiers. Input costs $1.50 per million tokens and output costs $7.50 per million tokens, and output includes thinking tokens, a detail that affects budgets because Hydrogen defaults to the `high` thinking level. Implicit caching runs automatically on prompts of at least 4,096 tokens and bills cached input at $0.075 per million tokens, a 95% discount. Explicit context caching, the `cachedContents` API, is not supported on this model, and Google's launch documentation does not list a Batch or Flex price for Hydrogen. Grounding with Google Search is billed per query.

The comparison with Gemini 3.8 Flash depends on the date. Gemini 3.8 Flash currently sells at an introductory $0.75 input and $3.75 output through December 31, 2026, and its price rises to $1.50 and $7.50 on January 1, 2027. Hydrogen therefore costs twice as much as 3.8 Flash for the rest of 2026 and exactly the same per input and output token from January, with a cheaper cache rate than 3.8 Flash will have at that point. Gemini 4 Argon, which Hydrogen beats on Google's reported Intelligence Index score, costs $2 and $10 until November 30 and doubles to $4 and $20 on December 1.

........

Model	Input	Output	Cached input	Batch	Price valid
Gemini 4 Hydrogen	$1.50	$7.50	$0.075	Not listed	From Nov 12, 2026
Gemini 3.8 Flash (intro)	$0.75	$3.75	$0.075	$0.375 / $1.875	Through Dec 31, 2026
Gemini 3.8 Flash	$1.50	$7.50	$0.15	$0.75 / $3.75	From Jan 1, 2027
Gemini 4 Argon (intro)	$2.00	$10.00	$0.10	$1.00 / $5.00	Through Nov 30, 2026
Gemini 4 Argon	$4.00	$20.00	$0.20	$2.00 / $10.00	From Dec 1, 2026
GPT-6.1 Terra	$1.00	$5.00	$0.05	$0.50 / $2.50	Current
Claude Sonnet 5.5	$2.00	$10.00	$0.10	$1.00 / $5.00	Current

........

All prices are in USD per million tokens at the standard tier, as listed on the official Google AI Studio pricing page, Google's Gemini 4 Hydrogen developer guide, the OpenAI API pricing page and the Claude platform pricing page on November 12, 2026. The Gemini 4 Hydrogen row comes from Google's launch post and developer guide, because the AI Studio pricing table had not yet added the model when it was checked. For Claude Sonnet 5.5, the cached input figure is the cache-hit price; writing to the cache costs $2.50 per million tokens for the five-minute cache.

··········

WHAT GEMINI 4 HYDROGEN COSTS FOR A SUPPORT CHATBOT AND A CODING AGENT.

Two workloads show where Hydrogen lands in practice. The first is a customer-support chatbot handling 10,000 requests a day, each with a 6,000-token system prompt and knowledge block that is served from cache, 2,000 new input tokens and 600 output tokens. On Hydrogen, one request costs $0.003 for the new input, $0.00045 for the cached input and $0.0045 for the output, or $0.00795 in total. That is $79.50 a day and about $2,385 over 30 days. The same traffic costs $1,260 a month on Gemini 3.8 Flash at its introductory price, $2,520 a month on 3.8 Flash from January, $3,180 on Gemini 4 Argon at its introductory price and $6,360 on Argon from December 1.

The second workload is a coding-agent session of 50 model calls, each with 80,000 input tokens of which 85% are cache hits, and 4,000 output tokens including thinking. Hydrogen bills that session at about $2.66. Gemini 3.8 Flash costs $1.46 at its introductory price and $2.91 from January, Argon costs $3.54 until November 30 and $7.08 afterwards, OpenAI's GPT-6.1 Terra comes to $1.77 and Anthropic's Claude Sonnet 5.5 to $3.54.

........

Model	Support chatbot, per month	Coding-agent session
Gemini 4 Hydrogen	$2,385	$2.66
Gemini 3.8 Flash (intro, to Dec 31)	$1,260	$1.46
Gemini 3.8 Flash (from Jan 1, 2027)	$2,520	$2.91
Gemini 4 Argon (intro, to Nov 30)	$3,180	$3.54
Gemini 4 Argon (from Dec 1)	$6,360	$7.08
GPT-6.1 Terra	$1,590	$1.77
Claude Sonnet 5.5	$3,180	$3.54

........

These figures assume identical token counts on every model, which will not hold in production. Hydrogen's default `high` thinking level can generate more output tokens than 3.8 Flash at its old `medium` default, while Google says partner tests saw Hydrogen use 12% to 23% fewer tokens than 3.8 Flash on complex graph, SQL and data-heavy enterprise workflows. Teams that run high volumes of short chat turns will pay less on 3.8 Flash until the end of December, so the economic case for switching now rests on quality and speed. Teams currently on Argon for routine agent work have the clearest saving, because Hydrogen undercuts Argon's introductory price and costs well under half of Argon's price from December 1. Developers comparing plans across Google's lineup can find the current figures in our guide to Gemini API pricing for Argon, Carbon and Decision 1.

··········

GEMINI 4 HYDROGEN BENCHMARKS AND SPEED CLAIMS.

Google's benchmark list centres on agentic and coding work. On Terminal-Bench 4.1, Hydrogen scores 57%, against 18.2% for Gemini 3.8 Flash. On Humanity's Last Exam in its Diamond subset, Google reports 58.8% and calls it the best result in Hydrogen's class. On Vals benchmarks, Google says Hydrogen ranks first overall on the Tax Agent Bench with 52.3% and ties for first on the Legal Research Bench with 55.8%. It reports 53.4% on Google's Multimodal Knowledge Agent evaluation, 71.6% on Agents' Last Exam v1.1, a fifth-place ranking on arena.ai's live Agent Arena, and a class-leading result on AutomationBench-AA.

........

Benchmark (as reported by Google)	Gemini 4 Hydrogen	Comparison
Artificial Analysis Intelligence Index	51 at high, 54 at xhigh	Gemini 4 Argon: 50
Terminal-Bench 4.1	57%	Gemini 3.8 Flash: 18.2%
HLE-Diamond	58.8%	Leads its class (per Google)
Vals Tax Agent Bench	52.3%	Ranked #1 overall
Vals Legal Research Bench	55.8%	Tied for #1
Multimodal Knowledge Agent	53.4%	Not specified
Agents' Last Exam v1.1	71.6%	Not specified
Agent Arena (arena.ai)	Ranked #5	Live leaderboard

........

The speed claims are relative to Gemini 3.8 Flash. Google says Hydrogen completes tasks 2.5 times faster at its default `high` setting, and that partner tests measured speedups between 2x and 2.8x. In Antigravity's Orchestrator mode, where Carbon coordinates and Hydrogen runs worker and verifier agents, Google says Terminal-Bench 4.0 runs reached task success close to Carbon alone while using 24% less quota and finishing 19% faster on average. Google also says Hydrogen produces Sheets financial-modeling outputs more than 30% faster than previous models. Most of these comparisons are against Google's own models; the launch post says Hydrogen matches or beats competing models in its class but does not name those competitors for most results, and the figures have not yet been reproduced independently.

··········

GEMINI 4 HYDROGEN API CHANGES FOR DEVELOPERS MIGRATING FROM GEMINI 3.8 FLASH.

Moving from Gemini 3.8 Flash requires changes beyond the model ID. Gemini 3.8 Flash used a 1-million-token input budget plus a separate 64K output budget, while Hydrogen shares one 1,048,576-token window across input, thinking and output. A synchronous call can return up to 65,536 tokens per hop, which is also the default cap when `max_output_tokens` is not set. Longer outputs, up to the remaining window, require background execution with `background: true` and a `max_output_tokens` value above 65,536, or client-side continuation.

Several parameters change behaviour. Thinking is controlled through `thinking_level` in the generation config, with `low`, `medium`, `high` and `xhigh` accepted; the `minimal` level that some earlier models supported is rejected with an error. The sampling parameters `temperature`, `top_p` and `top_k` are removed, and the server ignores custom values. Explicit caching through `cachedContents` and `caches.create` is not supported, so requests that carry a `cached_content` field must be cleaned up, and implicit caching takes over automatically above 4,096 prompt tokens. For multi-turn and tool-calling sessions, Google recommends `previous_interaction_id`, or returning prior model turns and thought signatures unchanged when the client manages the history. Tools should be declared natively or through structured outputs, because prompting for raw-text JSON tool calls without declared tools can return `MALFORMED_FUNCTION_CALL`.

Google's developer guide gives this minimal Python example for the new model:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-hydrogen",
    input="Build an interactive single-file HTML token bucket rate-limiter simulator with live sliders for burst size and request traffic."
)

print(interaction.output_text)

Teams coming from Gemini 4 Carbon or Argon need no request-schema changes. The differences there are the default thinking level, which is `medium` on Carbon and Argon and `high` on Hydrogen, the extra `xhigh` level, and a larger non-streaming output allowance of 65,536 tokens against 32,768 on Carbon and Argon over REST. Readers following the rest of the family can compare these limits with our coverage of Gemini 4 Carbon opening to API developers at $2/$10 and the Gemini 4 Argon global launch.

··········

GEMINI 4 HYDROGEN SAFETY RESULTS AND AGENT BEHAVIOUR.

Google pairs the launch with a set of agent-safety figures, all measured against Gemini 3.8 Flash unless stated otherwise. The company says Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, cuts misaligned actions and task shortcutting across tool-use and coding evaluations by a factor of 12, and is up to two times less likely to bypass permission boundaries when a task is impossible or blocked. It reports a 96.7% honesty rate on an internal agent-honesty evaluation and 99.9% chain-of-thought legibility, with no deceptive reasoning edits found across more than 200,000 audited turns.

On prompt injection, Google says Hydrogen is more than 2.3 times more resilient on the Gray Swan indirect prompt injection benchmark. It also says false refusals on HealthBench Pro stay below 0.4%, and that the model makes more than five times fewer fabricated claims than 3.8 Flash. These are vendor-reported results, several of them on internal evaluations, and Google has published a model card alongside the announcement for the full methodology.

··········

GEMINI 4 HYDROGEN: CONFIRMED FACTS AND OPEN QUESTIONS.

The confirmed facts are concrete. Gemini 4 Hydrogen launched on November 12, 2026, as `gemini-4-hydrogen`, at $1.50 per million input tokens and $7.50 per million output tokens with $0.075 cached input, flat across a shared 1,048,576-token window. It has four thinking levels with `high` as the default, supports implicit caching only, drops sampling parameters, and is available on the free and paid tiers of the Gemini API, in Gemini for business, in Antigravity for AI Pro and Ultra subscribers, in the Gemini app and in Google Search worldwide.

Several points remain open. Google has not published a Batch or Flex price for Hydrogen, the model was missing from the AI Studio pricing table on launch day, and the free-tier rate limits were referred to a separate page without figures in the launch materials. Vertex AI availability, the Gemini CLI and Google AI Plus were not mentioned. The benchmark and safety results are Google's own and mostly compare Hydrogen with Google models. On price, Hydrogen undercuts Gemini 4 Argon now and by a wide margin from December 1, matches Gemini 3.8 Flash's post-January rates, and remains above OpenAI's GPT-6.1 Terra, so the decision for most teams will come down to whether Google's speed and agent-accuracy gains hold up on their own workloads.

··········

·····

DATA STUDIOS

·····

datastudios.org