Hands-on with Gemini 4 Argon: top professional-work benchmarks, but poor quota efficiency and front-end experience

I'm a little lost for words.

Tonight, Gemini 4 Argon finally became generally available.

It had actually already been announced once on 10/1, the day after OpenAI's DevDay.

But that was essentially just an announcement. Hardly anyone could use it: access was limited to a cybersecurity program called Fairwind, mainly for defenders in government, critical infrastructure, and security companies.

After holding it back for two weeks, Google finally chose today to announce the launch.

For more than a month, new large models have been arriving in relentless waves. Here is a quick timeline.

9/4: GPT-6 Astra.

9/22: Grok 4.7.

9/23: Claude Opus 5.5, GPT-6 Sol, and Luna—three releases in one night.

9/30: OpenAI DevDay and GPT-6.1 Sol.

10/1: Gemini 4 Argon announced.

10/13: Meta's first frontier model, Muse Wonder.

10/14: Gemini 4 Argon becomes generally available.

That is pretty much all the big overseas players accounted for. At the model level, though, Anthropic is still indisputably in tier T0, with OpenAI close behind, Meta and Google chasing hard, and Grok still playing in the mud.

I had not paid for Gemini in a long time. Tonight I handed over another 100 US dollars to upgrade to Ultra and try Gemini 4 Argon.

After several hours of testing, my personal view is that it will still struggle to challenge Claude Opus 5.5 and GPT-6 Astra.

Let's go through it point by point.

One. Google's past half-year and more

Before discussing Argon, I think we need to talk about Google's past half-year or so.

That will make the mission of this model a little easier to understand…

Go back to month 11 last year, when Gemini 3 Pro launched. At the time, it was arguably one of the best models in the world.

Then Google entered a very surreal stretch.

It released Gemini 3.1 Pro in month 2, then promised at I/O in month 5 to release Gemini 3.5 Pro within a month.

But it kept being delayed. Word was that 3.5 Pro was several months behind schedule, partly because its coding capabilities had missed internal targets. The pressure from Anthropic and OpenAI was simply too great.

Eventually, after all those delays, it was scrapped internally…

Then, on 8/5, Google's AI organization underwent a major reshuffle.

Hassabis stepped down as Google DeepMind CEO to become DeepMind chairman and Alphabet chief scientist, with Koray taking over.

That same day, the legendary Jeff, affectionately called “Jiefu” in Chinese, announced he was leaving after 27 years at Google to start a company.

Jeff joined Google in 1999, when it was still a small company.

His name is behind foundational technologies that later supported all of Google: MapReduce, BigTable, and Spanner. Much of that work was pair-programmed by him and his longtime collaborator Sanjay, sitting together at a single computer.

In 2011, he co-founded Google Brain with Andrew Ng and others, and later led the team that built TensorFlow.

After Google Brain and DeepMind merged, he became Google DeepMind's chief scientist and one of Gemini's technical leads.

Google later created a new highest job level specifically for him and Sanjay: level 11.

There is even a programmer meme, “Jeff Dean Facts,” that has circulated for almost 20 years, inventing jokes to celebrate his abilities.

For example: “Compilers never warn Jeff; Jeff warns compilers.” Or: “Light originally traveled at only 35 miles per hour in a vacuum, until Jeff Dean spent a weekend optimizing physics.” That sort of thing…

And now this almost godlike figure has left.

Sanjay left with him.

Two days later, SemiAnalysis published an article titled:

Gemini is Cooked but GCP is Cooking.

In plain language: Gemini is finished, but Google Cloud is taking off.

Google has been selling huge amounts of TPU compute to rivals such as Anthropic. Its own models have faltered, while its pick-and-shovel business keeps growing.

For much of this half-year and more, Google has survived on the Flash series, from 3.5 Flash through 3.8 Flash on 9/2. Imagine OpenAI no longer being able to produce Astra and releasing Luna every day, or Anthropic unable to make Fable and releasing nothing but Haiku.

Tell me that isn't surreal…

So Argon is truly not just another model release for Google.

It is Google's last stand.

Two. The basics

As usual, let's start by summarizing the basic information.

Gemini 4 Argon is now live, and Pro subscribers paying 20 US dollars can use it in the Gemini app.

But to use it in Antigravity—Google's answer to Claude Code and Codex—you must subscribe to Ultra for 100 US dollars.

That is a real pain…

Unsurprisingly, it drew criticism.

There are three reasoning levels: low, medium, and high. The Chinese translation makes me wonder whether anyone on the team understands Chinese.

As is traditional for Gemini, this is a multimodal model, accepting nearly every kind of input: video, audio, images, text, and more.

Its knowledge cutoff is 2026-3, a little earlier than those of GPT-6 Astra and Claude Opus 5.5.

There is a small detail about the context window.

At the announcement on 9/30, Google said the output limit would jump from the previous generation's 6.4 × ten thousand tokens to 100 × ten thousand.

Today's wording, however, is “a combined input and output total of 100 × ten thousand.” Input, thinking, and output share that allowance, alongside a so-called continuation protocol.

This means Argon can still produce only about 6.4 × ten thousand tokens in one request, much like its predecessor. When it reaches the limit, it does not simply stop: it returns an encrypted continuation token recording where it left off. The developer then resends the original request unchanged with that token, and it continues writing.

It keeps going in segments this way, up to 100 × ten thousand…

GPT-6 Astra can natively output as much as 128K in one go, which is actually a little more than Gemini 4 Argon…

Next, pricing.

During the current promotion, input costs 2 US dollars and output 10 US dollars—the same prices as GPT-6.1 Sol and Claude Sonnet 5.5.

The promotion ends on 11/30, however. After that, pricing becomes 4 US dollars for input and 20 US dollars for output—exactly the same as Claude Opus 5.5.

For agents, though, caching is where much of the cost really lies. Here is a comparison.

Gemini 4 Argon: 0.1 US dollars during the promotion, then 0.2 US dollars after 12/1.

GPT-6.1 Sol: 0.1 US dollars.

Claude Sonnet 5.5: 0.1 US dollars, just reduced from 0.2 US dollars on 10/7.

Claude Opus 5.5: 0.2 US dollars.

GPT-6 Astra: 1 US dollars.

Now for benchmarks. Let's start with Google's own results, which I think… deserve no more than a quick look.

That is roughly it for the basics.

Three. Built for knowledge work

Beyond a company's own victory narrative, I often find third-party evaluations a little more objective.

Starting with AA: an overall score of 53, tied with GPT-6 Astra.

AA's benchmark gives substantial weight to coding and agents. Claude Opus 5.5 remains firmly in first place, followed by Claude Sonnet 5.5.

There is another detail here worth particular attention.

In AA's overall evaluation, Argon costs about 1.99 US dollars per task on average, compared with just 0.72 US dollars for GPT-6.1 Sol Max.

Yet both currently have the same API unit prices: 2 US dollars for input and 10 US dollars for output.

The main reason is that, compared with GPT-6.1 Sol, Argon still thinks and outputs far too much—though it is somewhat better than those Claude models that love marathon thinking.

But Claude's models may think at length, yet the quality of their completed tasks is extraordinarily high. It is hard to argue with that.

Another interesting third-party evaluation is called Vals.

Think of Vals as an independent evaluation company focused on professional expertise in fields such as finance, law, tax, medicine, and programming, with questions designed to resemble real work as closely as possible.

This time, they ran Argon through 26 tests.

On the overall Vals Index, covering finance, law, tax, and similar areas, Argon reached the very top of the Pareto frontier.

It scored 68.9, first overall.

Sonnet 5.5 scored 67.04%, Opus 5.5 66.97%, and Fable 5.1 65.83%.

Vals also calculated the cost per run.

Argon cost 15.68 US dollars, Sonnet 5.5 21.34 US dollars, Fable 5.1 28.71 US dollars, and Opus 5.5 32.14 US dollars.

Capable and cheap—less than half the cost of Opus 5.5.

Looking at the details, finance is where it actually ranks first. In law, tax, medicine, and programming, it is generally among the leaders.

One area was a complete letdown: on Vals's computer-use test, CUA-bench, Argon scored just 4.83%, second from last among 8 models. Only Gemini 3.8 Flash ranked lower.

GPT-6 Astra came first with 19.17%.

Claude Opus 5.5 came second with 14.00%.

That is a substantial gap in this weak area.

Overall, then, this Gemini model does reasonably well on specialized professional work, but cannot help you operate a computer.

To summarize Argon's strengths, the order is roughly:

Knowledge work > scientific research > coding > agent capabilities.

Four. A poor user experience

With the model covered, let's talk about the experience of using it.

I will not even discuss Gemini's web version. Anyone who has used it knows how bad the tool calling and thinking experience is.

Let's go straight to Antigravity, Google's competitor to Codex and the Claude desktop app.

First, language support. Codex supported a Chinese interface from its very first day. Even the money-grubbing people behind Claude added Chinese support to Claude. And don't start telling me it is some sort of trap: I have used it for a week without trouble.

But Antigravity has been out for almost a year, hasn't it? It still has nowhere to set the language.

I briefly wondered whether I was an idiot. I even asked Codex to look, and it really is not there.

Then there is the absurd rate at which the usage allowance disappears.

When Gemini 3 Pro first became famous, it was largely for front-end capabilities that felt a generation ahead.

I still remember how it amazed me back then. It really was powerful.

But today, coding front ends has become much more sophisticated. Whether it is GPT-6 Astra's superb spatial and 3D abilities or Claude Opus 5.5's exceptional visual taste and animation timing, both now seem to have the edge over Gemini.

Tonight I had planned to run 20 multidimensional aesthetic tests that I had specifically collected from popular community examples.

Instead, I completed barely a handful of tasks before my allowance was empty…

For goodness' sake, I am paying 100 US dollars. I did fewer than 10 tasks in total: 5 front-end tests, 3 content-creation tests, and 1 reverse-engineering test.

None of these was a huge job involving multiple agents in parallel. Yet my entire 5-hour allowance was gone, along with 27% of my weekly allowance.

Believe me, GPT-6.1 Sol on a 100-US-dollar subscription would not burn through its allowance this absurdly.

And the front-end output was unbelievable. It is already 2026-10, and it gives me an expanded HTML file for me to copy myself???

Not even an extra sentence at the end.

Literally just an HTML file.

In the end, I had to click Review, choose to open it from its folder, and manually double-click to run it before I could see the result in a browser.

As for the front-end result itself, it looks all right until you compare it… Here is Gemini 4 Argon's version.

On its own, the animated 3D result looks pretty good, right?

But now let's look at GPT-6 Astra's.

In terms of detail, it is simply in another league…

I had prepared back-end tasks too, but the quota limit prevented me from testing any of them. These tasks came from my own production projects: real bugs that had required repeated rework. I eventually organized those cases into a small evaluation set containing dozens of tasks.

Unfortunately, there was no allowance left.

For content writing, Gemini 4 Argon took first place in Arena's blind tests.

But please note: those tests are actually in English.

For Chinese writing, though, frankly, I find the AI style overwhelming.

For example, I distilled my article from a few days ago about OpenAI's mathematics release into a prompt. I gave Gemini 4 Argon the outline, structure, and my main arguments, deliberately leaving out samples of my past writing style. I mainly wanted to test its ability to produce content directly from a basic framework.

I would not merely call the result unsatisfactory; it was painful to read. All of Gemini 3's old habits were still there: bizarre, overblown metaphors and an obsession with quotation marks. Take a look for yourself.

Especially the heading below: which “mountain peaks of the immortals” had it conquered?

When I saw it, I honestly lost my composure for a moment…

To be fair, the model surely has many other strengths, such as specialized professional work. But those are areas I do not understand well enough to judge.

In the areas I do understand, combined with the core product's user experience, I can say very clearly that my several hours of testing tonight were extremely disappointing.

Final thoughts

If most of your tasks resemble mine, I would not particularly recommend subscribing to Ultra after reading this article. If you really want to try it, start with Pro at 20 US dollars.

Finally, here are my main current monthly subscriptions for reference.

1. ChatGPT Pro: 100 US dollars, about 670 yuan

When GPT-6 Astra first came out, I was on the 200-US-dollar plan. I was still refactoring AIHOT and was going through roughly 1 200-US-dollar account per day. I temporarily added 2 more 200-US-dollar subscriptions; even with resets, that was only just enough.

Claude Opus 5.5 is now so strong at building products that I downgraded ChatGPT. The 100-US-dollar plan, together with the compensatory 62500 and reset passes, feels sufficient.

Here are the situations where I use it now:

Automating development around feedback and PRs for my own and company projects, automating project operations and monitoring, and so on. GPT-6.1 Sol is still very good at improving and fixing bugs.

I also use it for targeted back-end and hardware optimizations or protocol reverse engineering. Tonight, for example, I had it optimize the new company's gateway and AP settings to improve everyone's network stability.

For most people, though, I still think ChatGPT is the most well-rounded option. Chatting is completely free, there is an image-generation allowance, and GPT-6.1 Sol is enough for building small things.

2. Claude Max: 250 US dollars. I subscribed through Google Play, which added a 50-US-dollar channel fee, bringing it to about 1675 yuan.

Not much to add: Claude Opus 5.5 is basically the standout model right now. Apart from some back-end and hardware optimization and browser control, I find it hard for GPT-6.1 Sol or GPT-6 Astra to compete with Opus 5.5. My overall impression is that its allowance is about 5 times Codex's. After a week of intensive development, design, and animation work, my weekly allowance was only 96% used when it reset this afternoon; I still had not exhausted it.

Almost all my usage is Claude Opus 5.5. I barely use Fable 5.1.

3. X Ultra: 200 US dollars. I subscribed today. You can pay on the web with a Chinese credit card, for about 1338 yuan. I went through Google Play and paid an extra 60 US dollars—painful.

Musk's new subscription system launched today, and I signed up straight away. The first month is half price, at 100 US dollars. It bundles Grok, Grok Bot, Cursor, and X together. Grok 4.7 is weak, but I was still willing to buy it purely for Grok Bot. Right now, it is much more useful than ChatGPT's Dots, and it even lets you use Claude Opus 5.5 as the back-end model.

I am an X user anyway, and Musk offers a generous allowance: the 200-US-dollar Ultra plan gives 50x, while Claude and Codex both give 10x, so I think it is reasonable value.

But if ChatGPT improves Dots so it is less foolish, and Musk's Grok 4.8 or 5 disappoints again, I probably will not renew.

4. Suno Pro: 8 US dollars. I pay annually for a discount, making it 53 yuan per month.

It is in a class of its own for music. I do not use it often, but occasionally need it, so it is essential. I bought an annual plan, mainly shared with the video team.

5. LibTV Supreme: 1499 yuan per month, discounted because I pay annually.

I occasionally use it to generate videos, but it is mainly for our video team, whose volume is huge. There is no need to subscribe unless you work with video. Video AI is painfully expensive right now.

Those are basically the services I use most often myself.

Adding all the subscriptions and API charges together, I spend more than 6000 yuan a month on AI. It is expensive; I would love to subscribe to just one platform.

Seen that way, ChatGPT is still the most comprehensive all-rounder, accessible to everyone.

As for Gemini 4 Argon, given its current model and product performance, I will not be adding it to my subscription lineup. I do not think I will pay for it again.

That is about it.

I hope this helps~

That is all. Since you have read this far, if you enjoyed it, please like, mark as reading, and share. You can also star my account to get new posts promptly ⭐. Thank you for reading, and see you next time.

>/ Author: Kazike

>/ For submissions or tips, please email: [email protected]