Gemini 4 Hydrogen: Price, Benchmarks, How to Use It and What Breaks vs 3.8 Flash

Google has a new everyday workhorse. Gemini 4 Hydrogen launched on 12 November 2026 as the fast, cheaper member of the Gemini 4 family, and it takes over from Gemini 3.8 Flash in the API, the Gemini app and Google Search.

The headline claim is bold: on the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default setting, one point above Gemini 4 Argon, Google’s own frontier model from October. It costs a fraction of Argon’s post-promo price.

This guide covers everything in one place: specs, every benchmark with its source, real pricing maths against Flash, Claude and OpenAI, how to use it in the app and the API (with code), what breaks when you migrate, and who should switch now.

What Gemini 4 Hydrogen Is

Gemini 4 now has three models, shipped in just over six weeks. Argon (October) and Carbon (early November) are the deep-reasoning frontier models. Hydrogen is the fast tier that most people will actually touch every day.

Google DeepMind head Koray Kavukcuoglu calls it “our daily driver for agentic workflows” and says it is “built for rapid execution”. In plain terms, it is the successor to the Flash line, tuned for coding, tool calls and multi-step agent work rather than one-off chat answers.

Google also says Hydrogen makes more than 5x fewer fabricated claims than 3.8 Flash, edits code more surgically, and recovers better when a tool call fails or a search comes back empty. Those are vendor claims, so treat them as a direction, not a measured fact.

Specs at a Glance

These come from Google Cloud’s model page and Google’s launch post. Google has not published parameter counts or architecture details.

Note: “short-term GA” matters if you build on it. Google can retire this exact model ID with 45 days’ notice, so keep the model name in config, not hard-coded across your app.

Benchmarks: Every Number and Its Source

Most numbers below are vendor-reported by Google. The Intelligence Index and cost-per-task figures come from Artificial Analysis (AA), an independent lab, as reported by VentureBeat. I have not run these myself.

Careful: Google’s 57% is on Terminal-Bench 4.1, while AA’s 54% is on 4.0. They are different versions, so do not compare them directly.

One more independent detail: according to AA data cited by OrcaRouter, Hydrogen at xhigh outputs about 198 tokens per second vs 111 for 3.8 Flash. But its time to first token was around 41 seconds vs 17 seconds, because it thinks longer before answering. For chat, that can feel slower even though tasks finish sooner.

Pricing and Real Cost per Task

Hydrogen is priced like 3.8 Flash’s future standard rate, which kicks in on 1 January 2027. Until then, Flash is still half the price per token. But price per token is only half the story.

Because Hydrogen solves more tasks with fewer wasted steps, AA’s estimated cost per benchmark task is much lower than Flash’s:

Tip: xhigh did not beat high on Terminal-Bench 4.0 (both 54%) but costs about 70% more per task. Start at high, and drop to medium or low for simple classification, extraction or chat.

For a full cross-provider view, see my AI API pricing comparison and the Gemini API pricing and free tier guide.

Where You Can Use It (App, Search, API, Antigravity)

Hydrogen rolled out globally on 12 November. Here is who gets it, and where.

Use Hydrogen in the Gemini app

1. Open gemini.google.com or the Gemini app on Android or iPhone and sign in.
2. Tap the model dropdown at the top of the chat.
3. Pick Gemini 4 Hydrogen if it is listed for your plan. On the free plan, leave it on auto and Gemini will route suitable prompts to Hydrogen.
4. Connect Gmail, Drive or apps like Dropbox and Airtable in settings if you want it to pull context from them.

Not sure which plan you need? My Google AI Plus vs Pro vs Ultra guide lists limits per plan.

The Antigravity Orchestrator

Antigravity’s new Orchestrator switches between Carbon and Hydrogen in one session. Carbon plans and coordinates, while Hydrogen runs the worker and verifier agents in isolated workspaces.

On Terminal-Bench 4.0, Google says this combo got close to Carbon’s success rate while finishing about 19% faster and using about 24% less quota (vendor-reported). If you live in Antigravity, see my Antigravity swarm guide and Gemini CLI vs Antigravity CLI.

How to Call Hydrogen From Python, Step by Step

You need a Gemini API key from Google AI Studio and the google-genai SDK. If you are new to the SDK, start with my Gemini API Python tutorial.

1. Create a key at aistudio.google.com (Get API key) and set it: export GEMINI_API_KEY="your-key"
2. Install or upgrade the SDK: pip install -U google-genai
3. Run a first request with the model ID gemini-4-hydrogen (code below).
4. Set the thinking level per request. Use low or medium for cheap, fast jobs and keep high for coding and agents.
5. Remove any temperature, top_p, top_k, thinking_budget and cached_content settings left over from Flash.

from google import genai
from google.genai import types

client = genai.Client()  # reads GEMINI_API_KEY

resp = client.models.generate_content(
    model="gemini-4-hydrogen",
    contents="Write a Python function that dedupes a list of dicts by 'id'.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="medium"),
    ),
)
print(resp.text)
print(resp.usage_metadata)  # check thinking vs output tokens

Printing usage_metadata is worth doing on day one. Thinking tokens are billed as output at $7.50 per 1M, and the new default (high) uses more of them than Flash did.

Function calling works the same way as on earlier Gemini models; see my Gemini function calling tutorial for tool definitions and the loop.

Migrating From Gemini 3.8 Flash: What Breaks

Swapping the model name is not enough. Google’s migration notes and Cloud model page list these changes. Four of them fail silently, with no error.

Warning: a hard-coded max_output_tokens of a few thousand can now be eaten by thinking, leaving an empty or cut-off answer. This is the most likely “Hydrogen returns nothing” bug.

Google’s deprecation table lists no shutdown date for 3.8 Flash yet, according to OrcaRouter. So there is no rush for latency-sensitive chat apps, but the 1 January price change removes Flash’s cost advantage.

Hydrogen vs Flash vs Argon vs Claude vs OpenAI

Prices are standard list rates per 1M tokens as of 12 November 2026. Scores are AA Intelligence Index unless stated.

Data Studios modelled a 600K-token document request at $0.96 on Hydrogen vs $1.26 on Terra and $1.28 on Sonnet 5.5, because Hydrogen’s price stays flat over long prompts. For short coding sessions, Terra came out cheapest per session in the same model. These are list-price estimates that assume equal token counts.

For frontier-tier comparisons, read Fable 5.2 vs GPT-6 Astra vs Gemini 4 Carbon.

Limits, Safety Claims and What Is Still Unknown

Known limits

• Not the cheapest: roughly 15x the per-token price of Haiku 5.5, GPT-6 Luna and Muse Ember.
• Slow first token at high and xhigh, since it thinks before answering.
• No explicit caching at launch; cache TTL control arrives on 12 January 2027 per Google Cloud.
• Flex capacity is limited at peak demand.
• Short-term GA: the model ID can be retired with 45 days’ notice.
• Rate limits: the public docs only list Batch enqueued tokens (3M at Tier 1, 400M at Tier 2, 1B at Tier 3, same as 3.8 Flash). RPM and daily limits per tier are shown inside AI Studio, not in the docs. My rate limits guide explains how tiers unlock.

Google’s safety claims (vendor-reported)

• Up to 81% less overreach and 79% less under-service in agent tasks.
• 12x fewer misaligned actions and shortcuts in tool-use and coding evals.
• Up to 2x less likely to bypass permission boundaries when blocked.
• 96.7% on an internal agent honesty eval; 2.3x more resilient to indirect prompt injection on Gray Swan’s IPI benchmark.
• False refusals on HealthBench Pro under 0.4%.

Still unknown

• Parameter count, architecture and knowledge cutoff.
• Exact free-tier daily limits in the API and per-plan prompt caps in the Gemini app.
• Whether Gemini CLI defaults to Hydrogen, and when 3.8 Flash will be shut down.

Troubleshooting

For other Gemini error codes (429, 500, 503, safety blocks), see my Gemini API errors fix guide.

My Take

Hydrogen is the most important Gemini 4 release for normal developers, more than Argon or Carbon. A fast-tier model matching last month’s frontier model on AA’s index changes what “default model” means.

The catch is that it is not a drop-in Flash replacement. Ignored temperature, the shared context window and the high default will quietly change outputs and bills, so test before switching production traffic.

My advice: move agent, coding and long-document workloads now, since cost per finished task is far lower than Flash. Keep simple, latency-sensitive chat on 3.8 Flash until late December, then switch before the 1 January price rise.

FAQ

Is Gemini 4 Hydrogen free?

Yes in the Gemini app and Google Search, where free users get it through automatic model selection. In the API it is listed for all tiers in AI Studio, and paid use costs $1.50 input and $7.50 output per 1M tokens.

Does Hydrogen replace Gemini 3.8 Flash?

It is the new fast tier and the upgrade path from 3.8 Flash in the app, Search and API. Flash still works and has no announced shutdown date yet.

Is Hydrogen better than Gemini 4 Argon?

On the Artificial Analysis Intelligence Index it scores 51 at high vs Argon’s 50, and 54 at xhigh. Argon may still be better on the hardest long reasoning tasks, which is what Google positions it for.

What is the xhigh thinking level?

A new fourth reasoning setting above high. It raised the AA score from 51 to 54 but cost about $3.16 per task vs $1.87, and did not improve Terminal-Bench 4.0.

Why are my temperature settings not working?

Hydrogen ignores custom temperature, top_p and top_k values without raising an error. Control output with prompts, examples or structured output instead.

How big is the context window?

1,048,576 tokens, shared by your input, the model’s thinking and its output. A single foreground reply is capped at 65,536 tokens.

Can I use Hydrogen in Antigravity for free?

No. Individual Antigravity access to Hydrogen needs a Google AI Pro or Ultra plan, and the Orchestrator is rolling out to Ultra first.

Is Hydrogen cheaper than Claude Sonnet 5.5?

Per token, yes: $1.50 / $7.50 vs $2 / $10. Sonnet 5.5 still scored higher on Terminal-Bench 4.1 in Google’s own table (63.9% vs 57%).

Sources

• Google blog: “Introducing Gemini 4 Hydrogen”, Koray Kavukcuoglu, 12 November 2026 (blog.google)
• Google Cloud docs: Gemini 4 Hydrogen model page and developer guide (docs.cloud.google.com)
• Google AI Studio docs: rate limits and billing tiers (aistudio.google.com/docs/rate-limits)
• VentureBeat: “Google launches Gemini 4 Hydrogen”, Carl Franzen, 12 November 2026 (pricing tiers, Artificial Analysis data, Box eval)
• SiliconANGLE: “Google releases Gemini 4 Hydrogen, a lower-cost model said to outscore Argon”, 12 November 2026
• TestingCatalog and MakeUseOf launch coverage, 12 November 2026
• OrcaRouter: “Gemini 4 Hydrogen vs Gemini 3.8 Flash: what breaks” (migration changes, AA speed figures; note it resells model access)
• Data Studios: “Gemini 4 Hydrogen vs GPT-6.1 Terra vs Claude Sonnet 5.5” (modelled costs)