twimco.Let’s talk ↗︎
Menu
Field notes / Reviews

Gemini 4 Hydrogen review: the cheaper model now edges Argon on Google's chart. Move the routine steps before December 1

Google released a fast model today that it says edges past Argon on an independent index, at less than half of Argon's December price. If you were keeping Argon on the routine steps of an AI job, that plan just changed.

Editorial engraving of a bright window-wall studio in daylight: a woman in a pale blue shirt quickly slides a tall stack of blank index cards into a wire tray beside a small brass stopwatch, while at a separate table an older man in a sage sweater and reading glasses works slowly through one thick leather binder held shut with a coral clip, a laptop, calculator and notebook between them.

Gemini 4 Hydrogen, released November 12, 2026, is the model to test for the routine steps of an AI job: sorting, drafting, extraction, small code changes and test runs. It costs $1.50 per million input and $7.50 per million output tokens, against Argon's $4 and $20 from December 1, and Google says it scores slightly above Argon on the Artificial Analysis Intelligence Index. Keep Carbon or Argon on the hard step, and keep client data off the free tier.

This morning's advice already needs an update

At 9:30 this morning we published our Gemini 4 Carbon review. The advice was simple: pay for Carbon on the hardest step of a job, and keep Argon on the rest.

A few hours later Google released Gemini 4 Hydrogen, and "keep Argon on the rest" stopped being the obvious answer.

Here's why. Hydrogen is Google's fast, everyday Gemini 4 model. Google says it scores slightly above Argon on the Artificial Analysis Intelligence Index, 51 to 50 at its default setting. And it costs $1.50 per million input tokens and $7.50 per million output tokens. Argon costs $2 and $10 until November 30, then $4 and $20 from December 1.

So from December, the routine steps you were planning to leave on Argon could run on a model that costs a bit over a third as much per token. If that holds up on your own work, it's worth testing now, nineteen days before Argon's price doubles.

What Google released on November 12

Gemini 4 Hydrogen is rolling out globally today:

  • Developers: the Gemini API and Google AI Studio, on both the free and paid tiers. Paid Google AI Pro and Ultra subscribers also get it in Google Antigravity, Google's coding agent.
  • Businesses: Gemini for business, through the API or the model menu, and Gemini in Google Sheets for eligible business plans.
  • Everyone: Google Search for all users. In the Gemini app, you can pick Hydrogen from the model menu; free users get it when Google's auto selection chooses it for the prompt.

The facts a buyer needs, from Google's pricing page and launch post on November 12:

  • Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across the whole 1-million-token window. Thinking tokens are billed as output.
  • Free tier: yes, unlike Argon and Carbon. On the free tier, Google's pricing page says your data is used to improve its products. On the paid tier, it isn't.
  • Thinking levels: four, low, medium, high (the default) and a new xhigh.
  • Context: 1 million tokens in, and up to 1 million tokens out.
  • Caching: implicit only. Explicit context caching isn't supported.

Google also introduced Orchestrator in Antigravity, which switches between Carbon and Hydrogen inside one coding session on its own. That's the same split we recommended this morning, now built into Google's own tool.

Decision guide

Hydrogen and Argon, as Google prices and describes them

ConsiderationGemini 4 HydrogenGemini 4 Argon
Google's suggested useFast everyday coding and agentic work; the execution model beside Carbon or ArgonDeep, long-horizon reasoning on hard problems
Price per million tokens to Nov 30, 2026$1.50 input, $7.50 output$2 input, $10 output
Price from Dec 1, 2026$1.50 input, $7.50 output (no change announced)$4 input, $20 output
Free tierYes; free-tier data used to improve Google's productsNot available
Thinking levelsLow, medium, high (default), xhighMedium is the default; see Google's Argon guide
Source: Google AI Studio Gemini Developer API pricing page and Google's November 12, 2026 Hydrogen launch post; prices as published on November 12, 2026.
The cheap model is no longer the weak model, so the real question is which steps still need the expensive one.

Who it's for

Teams that already have AI doing a lot of ordinary work and are watching the bill. Sorting incoming requests. Pulling fields out of documents. Drafting replies. Writing the small code change and running the tests. Summarizing a call. Those steps run thousands of times a month, and they're where a lower per-token price actually shows up on an invoice.

It's also for teams still on Gemini 3.8 Flash who have been hitting its limits. Google pitches Hydrogen as the upgrade from Flash, and the comparisons it publishes are mostly against Flash.

It's less of a change if your AI use is one hard job a week. If the work is root-cause debugging or a contract set that contradicts itself, Carbon is still the model Google points you to, and our Carbon review still stands for that step.

Where it looks strong

The price gap with Argon. This is the part you can bank on, because it comes from Google's pricing page, not a benchmark. Per token, Hydrogen is 75% of Argon's November price and 37.5% of Argon's December price.

Speed and accuracy against Flash. Google says Hydrogen finishes tasks 2.5 times faster than 3.8 Flash at its default setting, more than triples Flash's Terminal-Bench 4.1 score (57% against 18.2%), and cuts fabricated claims by more than five times. Fewer made-up answers matters more than speed for anything a customer reads.

Handling failure. Google says it "adapts cleanly when a tool call fails or a search returns empty." In real agent work, that's the moment most cheap models go wrong: an empty result turns into a confident invented one.

Pairing with a bigger model. On Terminal-Bench 4.0, Google says Orchestrator got close to Carbon's task success while using 24% less quota and finishing 19% faster. Again, Google's numbers, but they test exactly the setup most teams should be building: a strong model where it counts and a fast one everywhere else.

Where it falls short

Every number is Google's. The index score, the speed claims and the Orchestrator results all come from Google or its partners. None of them used your tickets, documents or codebase. A one-point lead over Argon on one index is not a promise that Hydrogen matches Argon on your hardest step.

Flash is still cheaper until January. Gemini 3.8 Flash costs $0.75 and $3.75 through December 31, half of Hydrogen's price. On January 1 it rises to $1.50 and $7.50, the same as Hydrogen. If a step works fine on Flash today, there's no rush to move it before the new year. Plan to test it before then.

The free tier isn't for client data. It's tempting to prototype on the free tier. Google's pricing page says free-tier data is used to improve its products, so keep customer records, contracts and anything under an NDA on the paid tier from the first test.

Default thinking is high. Hydrogen defaults to its high thinking level, and every thinking token is billed at the output rate. Many routine steps will pass at low or medium. Test those first, or the price advantage leaks away in thinking you didn't need.

Grounding costs extra. If a step uses Google Search grounding, the first 5,000 searches a month are free across Gemini 3 and newer models, then it's $14 per 1,000. A single request can trigger several searches.

How the work moves

Moving one step from Argon to Hydrogen

  1. Pick the step

    One routine step Argon handles today, such as sorting requests or drafting a reply.

  2. Pull fifty real cases

    Use real inputs on the paid tier, with the answers a person already approved.

  3. Run three levels

    Run Hydrogen at low, medium and high thinking and log the result and cost of each run.

  4. Price at December rates

    Compare the lowest passing level against Argon at $4 and $20 per million tokens.

  5. Switch or record why not

    Move the step before December 1 if it passes; otherwise note what failed and keep it on Argon.

Illustrative test plan for a single workflow step; not a measured result.

If you care about how Google Search sees your site

Hydrogen is now running Google Search for everyone. Google called it "the biggest model upgrade to Search yet." Google says Search now handles long, multi-part questions better, gives more visual answers, and has upgraded its query fan-out (the extra searches it runs behind one question) to connect people "to a broader range of helpful links than before."

Nobody outside Google can say yet what that does to any one site's traffic. What it does change is the kind of question that reaches you: longer, more specific, with several parts. A page that answers one narrow part of that question clearly has a better chance of being one of those links than a page that answers everything vaguely. Watch your Search Console queries over the next few weeks and look for longer ones. Our note on AI Overview impressions without clicks covers how to report on that honestly.

Plain verdict

Gemini 4 Hydrogen is the model to test for the routine steps of an AI job, and the date that matters is December 1, when Argon's price doubles. It doesn't replace Carbon on the hard step, and the benchmark lead over Argon is small and Google's own. But the price gap is real and published.

Here's the test we'd run this month. List every step you currently send to Argon. For each one, take fifty real cases, run them on Hydrogen at low, medium and high thinking, and record the result and the cost of every run. Use the lowest level that meets your quality bar, then price it against Argon at December's rate. Move the steps that pass before December 1. Leave the ones that don't on Argon or Carbon, and write down why. Our note on setting an acceptance gate before a model takes the job walks through the scoring.

If you're building an agent and want help deciding which steps deserve which model, that's our AI applications and agents work. Talk to us.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎