📑 Table of contents

Google launches Gemini 4 Hydrogen: 51 on the Artificial Analysis index, 2.5x faster than 3.8 Flash, $1.50/M input tokens

Actu IA 🟢 Beginner ⏱️ 15 min read 📅 2026-11-13

Google launches Gemini 4 Hydrogen: 51 on the Artificial Analysis index, 2.5x faster than 3.8 Flash, $1.50/M input tokens

🔎 A "fast" model that makes the model above it obsolete in four weeks

Four weeks. That's all the advantage of Gemini 4 Argon lasted. On November 13, 2026, Google DeepMind launched Gemini 4 Hydrogen, a model built for code and agentic work that scores 51 on the Artificial Analysis Intelligence Index in thinking high by default — versus 50 for Argon, released a month earlier.

The nuance that changes everything: Hydrogen is not yet another premium model. It takes over from Gemini 3.8 Flash, Google's "fast" workhorse, with 2.5x faster task execution and a flat rate of $1.50/M input tokens, from the first to the millionth context token.

Better still: it doesn't arrive alone. According to e-acc.ai and aibriefs.news, the model is available from day 1 in Cursor, OpenClaw, AI Studio, the Gemini API, and Google Search. The llmgateway.io timeline confirms a global rollout across all Google surfaces, Gemini app included.

My blunt take: this is the most strategic release in the Gemini 4 family. Not the most powerful — Carbon keeps that title — but the one that tackles the real bottleneck of agentic work: cost and latency.


The Essentials

  • 51 on the Artificial Analysis Intelligence Index in default high thinking, 54 in xhigh — ahead of Gemini 4 Argon (50), one month after its release.
  • 2.5x faster than Gemini 3.8 Flash on task execution, with an index of 51 vs 31.
  • Flat pricing across the full 1M context: $1.50/M input tokens, $7.50/M output, $0.075/M on cache hits.
  • Credible agentic performance: 57% on Terminal-Bench 4.1 (vs 18.2% for 3.8 Flash), #1 on the Vals Tax Agent Bench, overaction reduced by up to 81%.
  • Antigravity's Orchestrator: dynamic switching between Carbon and Hydrogen within a single session — nearly Carbon's success rate with 24% less quota and 19% faster on Terminal-Bench 4.0.
  • Day 1 everywhere: AI Studio, Gemini API, Gemini app, Search, Antigravity, Cursor, and OpenClaw.

Hydrogen is available through the official API, but also in third-party tools from day one. Where to start:

Tool Main use Price (November 2026) Best for
Google AI Studio Test Hydrogen, compare with Argon and Carbon Free (limited quota) First tries and homemade benchmarks
Gemini API Integrate Hydrogen in production $1.50/M input, $7.50/M output Building agents
Cursor Coding agent with Hydrogen from day 1 ~$20/month (check cursor.com) Developers
OpenClaw Open source agent hooked up to Hydrogen Free (open source, + API cost) Personal automation
Google Antigravity Orchestrator that routes Carbon/Hydrogen AI Pro/Ultra subscription required Automatic routing
Hostinger Host the app that consumes the API from ~$2.99/month (check hostinger.com) Deploy a prototype online

51 on the Artificial Analysis index: what is this score really worth?

Direct answer: Hydrogen beats Gemini 4 Argon on "raw" intelligence, four weeks after Argon's release. That's the major fact of this launch — and it comes from an independent source, not a press release.

The Artificial Analysis Intelligence Index aggregates a series of reasoning benchmarks into a single score. It's the industry's compass for comparing models' gray matter, beyond the marketing slides. Hydrogen scores 51 on it in thinking high — the default level — and 54 in xhigh, the top tier of the four available levels (low, medium, high, xhigh), as The Decoder details.

Google confirms the figure in its official announcement: 51 versus 50 for Argon, 54 in xhigh. The kicker? Hydrogen is positioned as the "fast" model in the lineup. Argon, for its part, was never meant to be the cheapest.

Artificial Analysis' independent evaluation goes further: in xhigh, Hydrogen plays roughly in Claude Sonnet 5.5's league. But let's keep a cool head. On Terminal-Bench 4.1, Hydrogen tops out at 57% while Sonnet 5.5 reaches 63.9%. The composite index isn't an agentic benchmark, and the two tell very different stories.

So what's remarkable isn't the 51 in itself. It's the 51 in a model priced at $1.50/M for input. That level of intelligence at that price point — that's the real news of November 2026.


Pricing: the flat pricing that (finally) simplifies the agent bill

Direct answer: $1.50/M input tokens, $7.50/M output tokens, with no context tiering up to 1M. A single pricing grid, from the first token to the millionth.

Item Rate (November 2026)
Input (context up to 1M) $1.50/M tokens
Output $7.50/M tokens
Cache hits $0.075/M tokens
Cost per agentic task (xhigh, Artificial Analysis eval) ~$3.16

Two details deserve your attention. First, the cache: at $0.075/M, it costs 20x less than standard input. For an agent that resends the same system context at every iteration, this is the line item that slashes the bill. Next, the context: 1M tokens for both input and output, according to The Decoder — rare on the market, and billed at the same price as 10K tokens.

The cost per task tells the same story. In xhigh, Hydrogen runs around $3.16/task, roughly a third of the cost of Claude Sonnet 5.5 (max) for a comparable level of intelligence, according to Artificial Analysis. But beware: that's also 2.5x the cost of GPT-6.1 Sol (max). Hydrogen isn't the cheapest model on the market — it has the best intelligence/price/latency ratio in its class.

Let's be honest about one thing: VentureBeat describes a "more expensive" Flash successor. Hydrogen costs more than 3.8 Flash. Google's "fast" segment is becoming less "cheap," and that's a deliberate choice: raising intelligence rather than defending the price floor. Rates to verify on ai.google.dev, they change every month.


Speed: 2.5x faster than 3.8 Flash, and it shows on agents

Direct answer: Hydrogen is 2.5x faster than Gemini 3.8 Flash at default thinking level, and its score follows the same slope — an index of 51 versus 31.

Let's put the lineage in perspective. In early 2026, Gemini 3.5 Flash had already shaken up the fast-model hierarchy by beating Opus 4.7 and GPT-5.5 on agentic benchmarks. 3.8 Flash then took over on speed, but remained laughable at agentic work: 18.2% on Terminal-Bench 4.1. A fast model, but one you didn't want to let drive.

Hydrogen fixes that flaw without sacrificing speed. VentureBeat reports on the early access test run at Box: the agent goes from 74% to 79% accuracy, in less than half the time, with roughly 12% fewer tokens.

Why it matters: an agent loops. Plan, execute, correct, execute again — twenty iterations for a serious task. Latency multiplies with every loop. A model that's 2.5x faster isn't a luxury: it's the difference between an agent you can use during the day and one you launch in the evening with your fingers crossed.


Agentic: Terminal-Bench, Vals, and the end of over-action

Direct answer: Hydrogen is Google's first "fast" model that's credible at serious agentic work, not just chat.

Numbers first. On Terminal-Bench 4.0, Hydrogen reaches 54% in both high and xhigh, versus 40.9% for Muse Ember, 32.8% for Claude Haiku 5.5, and 12.6% for GPT-6 Luna, according to Artificial Analysis. On version 4.1 of the benchmark, it climbs to 57% — still behind Claude Sonnet 5.5 (63.9%), but ahead of everything else in its segment.

On business benchmarks, it's worse (for the competition): No. 1 on Vals Tax Agent Bench at 52.3%, tied for No. 1 on Vals Legal Research Bench at 55.8%, class leader on HLE-Diamond at 58.8%. On Agent Arena, which is dedicated to open-ended agentic tasks, it ranks 5th. For a model at $1.50/M, that's unsettling.

The most interesting number doesn't come from a public benchmark: Google claims Hydrogen reduces agentic over-action by up to 81%, with a 96.7% honesty rate on its internal evaluation. Vendor figures, to be taken with a grain of salt — but the concept matters. Over-action is the hidden cost of agents: unnecessary actions, burned tokens, cascading errors. The Decoder adds that hallucinations would be cut fivefold compared to 3.8 Flash. A model that takes fewer actions also makes fewer things up.

If you're looking for a model for coding or driving tools, our selection of the best LLMs for coding includes this launch.


Antigravity's Orchestrator: Carbon and Hydrogen in the same session

Direct answer: the real product innovation of November 13 isn't Hydrogen on its own — it's Antigravity's Orchestrator, which dynamically switches between Carbon and Hydrogen over the course of a single session.

The principle: the Orchestrator assesses the complexity of each step and routes to the right model. Trivial task → Hydrogen. Gordian knot → Carbon. Result measured by Google on Terminal-Bench 4.0: nearly Carbon's success rate, with 24% less quota and 19% faster. For the details on the flagship, read our article on Gemini 4 Carbon, Google's most powerful model.

It's a direct answer to the question everyone has been asking for the past six months: "which model should I choose?". Google's answer: none of them — the system chooses. Until now, we've been cobbling together automatic routing with homegrown frameworks and cost heuristics. Google ships it inside the tool.

My one gripe: it's elegant lock-in. The more your pipeline depends on the Orchestrator, the higher your exit cost climbs. If that makes you break out in hives, stick with the Gemini API and build your own router — you'll lose a few quota points, but you'll gain independence.


Facing the competition: behind Sonnet 5.5, ahead of everything else in its segment

Direct answer: Hydrogen takes second place in agentic benchmarks, just behind Claude Sonnet 5.5, at a third of the cost per task — and far ahead of the other models in its price bracket.

Model Terminal-Bench 4.1 Terminal-Bench 4.0 Takeaway
Gemini 4 Hydrogen 57% 54% The reference point for intelligence-to-price ratio
Claude Sonnet 5.5 63.9% — Still ahead, at ~3x the cost per task
Gemini 3.8 Flash 18.2% — Predecessor, outclassed on every front
Muse Ember — 40.9% Strong performer in the fast segment
Claude Haiku 5.5 — — ...

The question "ChatGPT or Gemini?" is also posed differently since this launch: our ChatGPT vs Gemini comparison now starts from the premise that the choice comes down to the agentic ecosystem32.8% | Gap widened |
| GPT-6 Luna | — | 12.6% | Lagging behind on agentic tasks |

Three possible readings. The first: Claude Sonnet 5.5 remains the king of agentic tasks, period. If every point of success rate is worth money, pay for it. The second: GPT-6.1 Sol (max) remains cheaper per task, but Artificial Analysis places Hydrogen xhigh on par with Sonnet 5.5 — the quality/price ratio tips in Google's favor. The third: pressure is also coming from open-weight models, with releases like MiniMax H3 squeezing the margins of closed models from below.

The question "ChatGPT or Gemini?" is also posed differently since this launch: our ChatGPT vs Gemini comparison now starts from the premise that the choice comes down to the agentic ecosystem, not the chat. For an up-to-date overview, check out our roundup of the best LLMs of November 2026.

My verdict: for 80% of common agentic tasks, Hydrogen becomes the rational default choice. Sonnet 5.5 keeps the critical 20%.


How to use it today

Direct answer: Hydrogen is available on six Google surfaces and with two third-party partners, from day one — no waiting list.

On Google's side: AI Studio for free testing (limited quota), the Gemini API for production, Gemini for business for organizations, the Gemini app, Google Search, and Antigravity for AI Pro/Ultra subscribers. In the app's picker, Hydrogen inherits the label "daily help," alongside 3.5 Flash-Lite (quick responses), 4 Argon (code & analysis), and 4 Carbon (most complex tasks), according to 9to5Google. Positioning a model at index 51 as a "daily" tool says a lot about the speed at which Google is devaluing its own models.

On the third-party side: Cursor and OpenClaw integrate it from day one, which has become the standard for a credible launch. A model absent from coding agents on day one is a dead model.

To test via command line:

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-hydrogen:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -d '{"contents":[{"parts":[{"text":"Résume ce dépôt et propose un plan de refactorisation"}]}]}'

Zero budget? Our guides on free AI APIs and the best free LLMs cover alternatives without a credit card. And if you are building a product around the API, you will need to host the interface: Hostinger is more than enough for a prototype.


A busy November at Google: the agentic stack comes together

Direct answer: Hydrogen doesn't launch in a vacuum. It's the centerpiece of a November in which Google rolled out the entire agentic value chain.

Three launches in a few weeks. Gemini Decision 1 for evaluation — up to 1,000 scored questions per call, $0.10/M input tokens and zero output cost. Gemini 3.5 Transcribe for perception — an STT engine covering 85 languages at 2.6% WER, 70% faster than Chirp 3, that even strips out hesitations on the fly. And Gemini 3.5 Flash, earlier in the year, to lay the foundations of the fast segment.

The puzzle is easy to read: perception (Transcribe), execution (Hydrogen), orchestration (Antigravity), evaluation (Decision 1). Google no longer ships models — it ships an integrated agentic production chain. It's Apple's hardware strategy, applied to models.

My take: this is the real threat to the competition. Not the 51 on an index — indexes can be caught up with. A complete stack is hard to replicate.


❌ Common Mistakes

Mistake 1: comparing Hydrogen and Carbon as competitors

They're not playing in the same league. Carbon aims for maximum success rate, Hydrogen for the intelligence/price/latency ratio. The right question isn't "which one is better?" but "which one for which stage?". If you're hesitating, the Orchestrator decides for you.

Mistake 2: confusing the default score with the xhigh score

The 51 is obtained with thinking high, the default level. The 54 requires xhigh mode — slower and more expensive. Citing 54 while presenting Hydrogen as "fast" means lying to yourself about the real latency. Test both tiers on your workload before making a decision.

Mistake 3: ignoring the cache

At $0.075/M versus $1.50/M for standard input, the cache costs 20x less. Agents send the same context on every loop: without caching, you pay ten times for the same prompt. Enable it from the very first iteration, not after the first bill.

Mistake 4: believing that "2.5x faster" means "2.5x cheaper"

Speed and cost are two independent variables. Hydrogen is faster than 3.8 Flash, but also more expensive — VentureBeat explicitly describes a "pricier" successor. Your bill depends on the tokens consumed and the thinking tier, not on clock speed.


❓ Frequently Asked Questions

What is the exact price of Gemini 4 Hydrogen?

$1.50/M input tokens and $7.50/M output tokens, flat across the full context up to 1M, with cache hits at $0.075/M. Pricing as of November 2026 — verify on ai.google.dev before making any architecture decisions.

Does Gemini 4 Hydrogen replace Gemini 3.8 Flash?

Yes, it takes over in the fast segment — VentureBeat describes it as its successor: faster, smarter, and more expensive. 3.8 Flash remains available via the API for workloads where price matters more than intelligence.

Hydrogen or Carbon: which one should you choose?

Carbon for the hardest tasks where success rate is paramount, Hydrogen for everything else. In practice, Antigravity's Orchestrator switches between the two within a single session and achieves nearly Carbon's success rate while using 24% less quota.

Can you use Gemini 4 Hydrogen for free?

Yes, in AI Studio with a limited quota, and in the Gemini app depending on your subscription. Antigravity, however, requires an AI Pro or Ultra subscription. To test without a credit card, check out our selection of free AI APIs.

Should you migrate your agents to Hydrogen now?

If your agents run on 3.8 Flash, yes: 57% vs 18.2% on Terminal-Bench 4.1 speaks for itself. If you're on Claude Sonnet 5.5, test first — Hydrogen performs worse on pure success rate, but is potentially three times cheaper per task.


✅ Conclusion

With Gemini 4 Hydrogen, Google delivers a model with an index score of 51, 2.5x faster than 3.8 Flash and billed at $1.50/M input, in the hands of every coding agent from day 1 — the intelligence/price/latency ratio of the fast segment just got a new reference point. To see where it ranks against the rest of the market, check out our ranking of the best LLMs of November 2026.