Gemini 4 Hydrogen Is Google's Fast Tier, but the Discount Is Smaller Than It Looks

Gemini 4 Hydrogen Is Google's Fast Tier, but the Discount Is Smaller Than It Looks

6 min readNovember 14, 2026

Quick verdict

Google filled out the Gemini 4 line with a third model, and this one is about speed. Hydrogen sits below Carbon and Argon, runs at roughly 200 tokens per second, and Google says it finishes tasks 2.5x faster than Gemini 3.8 Flash. It has one genuinely standout number: an 8% hallucination rate on Artificial Analysis's Omniscience test, against 43% for Claude Sonnet 5.5 and 52% for GPT-6.1 Sol. It is also everywhere already, including a free tier in AI Studio and auto-selection in the free Gemini app. The catch is the price tag. At $1.50 input and $7.50 output per million tokens, it is only about 25% below Carbon, not the order-of-magnitude drop the "Flash" branding trains you to expect, and its raw coding scores still land behind Sonnet and Sol. So it is fast, honest, and widely available, but it is not the cheap workhorse the name suggests.

What actually shipped

Hydrogen is the speed model in the Gemini 4 family, sitting next to Carbon and Argon. Here is what Google led with:

  • It finishes tasks 2.5x faster than Gemini 3.8 Flash, running at roughly 200 tokens per second.
  • There are four thinking levels, low, medium, high, and a new xhigh, plus a 1M-token context window.
  • Pricing is $1.50 per million input tokens and $7.50 per million output, with a 95% discount on cached input.
  • On Google's own benchmarks it hits 57% on Terminal-Bench 4.0, more than 3x Gemini 3.8 Flash on Terminal-Bench 4.1, over 5x fewer fabricated claims, and the top spot on Vals Tax Agent Bench.
  • You can already use it on every Gemini API tier, including a free tier in AI Studio. It is the default in Search AI Mode, free Gemini app users get it by auto-selection, and Pro and Ultra subscribers can run it in Antigravity.

On the index Google cares about most, Hydrogen edges its bigger sibling: 51 versus Argon's 50 on the Artificial Analysis Intelligence Index at default thinking, climbing to 54 at the new xhigh level. That is a real result for a model meant to be the quick one rather than the smart one.

Where it lands against Claude and GPT

The independent read is more useful than the launch slides, and Artificial Analysis ran Hydrogen through its full suite. The profile is lopsided in an interesting way.

Metric (xhigh)Gemini 4 HydrogenClaude Sonnet 5.5 (max)GPT-6.1 Sol (max)
Intelligence Index545555
Hallucination rate (AA-Omniscience)8%43%52%
Terminal-Bench 4.054%64%56%
Cost per task$3.16~3x Hydrogen~0.4x Hydrogen

Read that top to bottom and you get the shape of the model. On general intelligence it is one point behind the two frontier flagships, which is close. On hallucination it is not close, it is in a different league, and for anything where a made-up fact costs you real money that is the number that matters. But on pure coding it slips: 54% on Terminal-Bench 4.0 at both high and xhigh, under Sonnet's 64% and Sol's 56%. And the cost math cuts both ways. Hydrogen runs at about a third of Sonnet's cost per task, but roughly 2.5x GPT-6.1 Sol's, so "cheap" depends entirely on which flagship you were going to use.

The third-party leaderboards fill in the rest. Hydrogen debuts at #5 on Agent Arena with a +14.8% net improvement across 6,500 sessions, and it is the fastest model in the top 10 at a 5.3-minute median per task. On Vals it sits #6 at 68.5%, two points behind Carbon while costing 38% of Carbon's price. On Cursor's own CursorBench it jumps to 54.3% from 3.8 Flash's 39.6%, a solid generational step for coding even if it does not top the chart.

The orchestrator mode is the part to watch

The most Admix-relevant thing in this launch is not Hydrogen alone, it is how Google wants you to use it. In Antigravity, Google is rolling out a multi-model mode where Carbon plans the work and Hydrogen writes and tests the code. Google says the pair reaches near-Carbon success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. That is the whole thesis behind running more than one model: a smart-but-slow model for planning, a fast-but-cheaper model for execution, and a better result than either on its own. If you already think in terms of picking the right model per task instead of paying one flagship to do everything, this is that idea shipping inside Google's own tool.

Where the pitch wobbles

Two of the sharpest voices in the space pushed back on launch day. Simon Willison pointed out that the "fast tier" framing hides a thin discount: Hydrogen is only about 25% cheaper than the Carbon and Argon token price, and those two models reportedly double in price on December 1, which quietly makes Hydrogen look better by comparison. Theo disputed the "built for speed" claim outright based on his own timings. And testers found rough edges: one watched xhigh burn 131,000 thinking tokens and about $1.08 without ever answering a question about a jersey number, a failure mode where the model reasons itself into a hole. Hydrogen also scored 1% on MazeBench, so spatial puzzles are not its thing.

None of that sinks the model. It means the honest summary is narrower than the launch: Hydrogen is a fast, unusually truthful model with a wide free rollout, priced closer to a mid-tier than a budget tier, with a runaway-reasoning risk you should cap with a lower thinking level for routine work.

Video: Gemini 4 Hydrogen in context

A rundown of where Hydrogen fits next to Carbon and the rest of Google's new releases, and what the speed and pricing actually look like in use.

FAQ

Is Gemini 4 Hydrogen free?

Partly. There is a free tier in Google AI Studio, it is auto-selected for free users in the Gemini app, and it is the default in Search AI Mode. Heavy API use runs $1.50/$7.50 per million tokens, with a 95% discount on cached input.

Is Hydrogen better than Claude or GPT for coding?

Not on raw coding benchmarks. It scores 54% on Terminal-Bench 4.0, behind Claude Sonnet 5.5 at 64% and GPT-6.1 Sol at 56%. It wins on speed and on hallucination rate, not on top-end code quality. If coding accuracy is the priority, see our best AI models for coding in 2026.

Why would I use Hydrogen over Gemini 4 Carbon?

Speed and price. Hydrogen runs far faster and costs about 38% of Carbon's price while landing within two points of it on the Vals Index. Carbon is the better choice when you need its top-end reasoning; Hydrogen is the better default for high-volume work.

What is the 95% cached-input discount?

If you send the same context repeatedly, like a long system prompt or a document, Google charges 5% of the normal input price for the cached portion. It mainly helps agents and chat apps that reuse the same prefix across many calls.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 300+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.

Try Admix for $1.99

Related articles