Gemini 4 Argon: Top Marks on Familiar Tests, Behind on New Coding Tests

Gemini 4 Argon is the most powerful model Google has released so far. It was announced on September 30 and opened to everyone on October 13. In the table Google published, it posts the top score on 13 of 19 tests, and against Claude Opus 5.5, the model that leads the independent rankings today, it wins 14 of 18.

I work with models from Anthropic, OpenAI and Google every day, and since the launch I have followed three kinds of evidence: Google's table, tests by independent organizations, and checks by developers who ran the same task on Argon and on Opus. I did not run Argon myself on client tasks. Throughout the article, I flag what is fact and what is my opinion.

What Gemini 4 Argon is, and why it took two weeks to open

Argon is the large model in the Gemini 4 family, built for the hardest reasoning. Alongside it, Google runs separate models: Gemini 3.8 Flash for speed, Nano Banana for images, Omni Flash and Gemini 3.8 TTS for voice, and Lyria for music. Until now, Google's strongest model was Gemini 3.1 Pro, and version 3.5 Pro, promised for June, was cancelled along the way.

On September 30, Google announced Argon but opened it only to a closed group of cyber defenders, under a program called Fairwind. On October 13, 13 days later, it opened to everyone:

Sources: Google's launch announcements (September 30 and October 13), the Gemini API price list and the model's developer page. If the term is new to you, we explained what a token is in a separate article.

Where you can use it today:

• In the Gemini app: for Google AI Pro and Ultra subscribers. Free users don't get it.
• In the API, the direct connection for developers and systems: paid tier only. There is no free tier.
• In Google's coding tool, Antigravity: Ultra subscribers only.

The official explanation for the delay is safety. Google wrote that a staged rollout is needed to release capabilities at this level safely, and it took part in a voluntary US government program that gets early access to models. In the model card, it states that it cannot rule out that Argon has crossed the critical threshold it set for cyber capabilities. What I think about the delay is in the opinion section.

One thing worth knowing before we go on: Argon itself outputs only text. In Google's demos, the images came from Nano Banana, the engine sounds from Omni Flash, the narration from Gemini 3.8 TTS and the music from Lyria. Argon orchestrates them and writes the code. Google's strength in images, video and design belongs to the whole family, not to one model.

Google's table: 13 wins out of 19

Google published a table of 19 tests, pitting Argon against three competitors: OpenAI's GPT-6 Astra, and Anthropic's Claude Fable 5.1 and Claude Opus 5.5. In each row the top score is marked in green, and a label flags rows where Google ran Argon itself:

On paper, it's a clear win. But read the methodology document Google published alongside the table, and three things change the picture.

First: some of the conditions favor Argon. In computer use, Argon's score is the best of three runs. On the command-line science test, it got 6 times the usual time limit. In video understanding, each model saw a different number of frames from the video. Each of these is written in the document; none of them is written in the table.

Second: two competitors are missing. Claude Sonnet 5.5 and GPT-6.1 Sol cost the same as Argon at launch pricing, $2 and $10 (for Sol, up to 272K input tokens), and neither is in the table. Sonnet 5.5 even scores higher on the Artificial Analysis index: 56 against 53. Anyone choosing a model by price should compare Argon with these two.

Third: where Argon loses. Out of four coding tests, it loses two: FrontierSWE, where the tasks are long and new, and Terminal-Bench. Google never wrote "the best model in the world," but it did write that Argon "outperforms other leading models on benchmarks across every domain." That part doesn't hold up when you look at code.

What the independent organizations say

A maker's table is a report card where the student picked which grades to include and whom to be compared against. To know where Argon really stands, you need organizations that run all the models under the same conditions. Here is the overall index from one of them:

On this index, Argon sits below Opus 5.5 and Sonnet 5.5, tied with GPT-6 Astra and Claude Fable 5.1, which came out in early September. That is a very good level, but not the top of the table.

In the other independent sources, the picture splits by type of work:

Sources: Arena (September 30), Vals AI (September 30), Epoch AI data covering CursorBench, FrontierCode and APEX-Agents (October 14), Artificial Analysis (October 14).

The pattern repeats: in writing and in business tasks with a checkable answer, Argon leads or is near the top. In code, interface building and professional tasks where experts judge the quality, it trails. Both lists include new and private tests, so the gap follows the type of work. Even so, the new coding tests stand out on the red list.

benchmaxxing: did Argon "study for the test"?

benchmaxxing is a nickname for a model trained to excel specifically on the well-known tests. Picture two students who both get a perfect score on the SAT. One learned the material; the other worked through every past exam. On the next test, with questions neither has seen, the difference will show. To be precise: this is not cheating. It's a decision about where to invest the training, and the result is a model that looks better on the table than at work.

On September 30, announcement day, Bloomberg reported that Google employees were saying exactly this about Argon: that it posts good test scores and does less well when you put it to work, especially on interface design. According to the report, two people familiar with the model said Gemini 4 "appears to be affected" by this process. Google said it is "inaccurate to say Gemini 4 lags in areas like coding," and published an internal figure: among its engineers, 35% more code written by AI agents and almost a third fewer changes rolled back.

The pelican test

Simon Willison's pelican test is a long-running unofficial test: "draw a pelican riding a bicycle," as code for an SVG file. Every new model gets it, and a good result spreads fast. Argon produced an excellent pelican, at a cost of 17 cents.

Willison himself does not claim that Google trained Argon on the pelican. He does note that Google is the only lab he knows of whose employees have put special effort into SVG drawings. In the Hacker News discussion, some commenters asked whether the test has been "burned." That is the fate of every famous test: once everyone knows it, it measures less.

The check I ran: which month is Argon in?

To test my sense that pelicans aren't everything, I took the four newest coding tests that include an Argon result and checked which models it beats and which it loses to. The date is when each model was released:

On two of the four tests, website building and FrontierCode, Argon sits below Claude Opus 5 from late July, at the level of models from mid-July. On FrontierCode it is even below Claude Fable 5 from June. In other words, in coding and interface design, Argon is where the frontier was about three months ago, and that matches what I see in the examples. On the other two tests it is closer, at September's level.

And what argues against the suspicion

Not all the evidence points the same way:

• Arena is a very hard test to train for. The questions are written fresh every day by people, and Argon is first there in text. On the other hand, Arena measures what people prefer to read, and a model can learn to write answers that are pleasant to pick. The researcher Teortaxes wrote that Argon may be "tuned to preference data," and that on tests that are harder to train for, its scores are impressive.
• On new tests that aren't about code, Argon is strong. It is first on APEX-Agents (analyst and consultant tasks) and on Blueprint-Bench 2 (spatial reasoning), and on Vending-Bench 2 (running a simulated business) it beats Opus 5.5, though not GPT-6 Astra.
• Leon Lin, who tested Argon before launch, wrote that it is "not as benchmaxxed as a lot of people think," and in the same post that it is weak at interfaces and animation, and that it "has no taste."

Leon Lin's post on X, October 13, 2026

My conclusion from the data: benchmaxxing doesn't explain the whole model. It mainly explains the gap between Google's coding claims and what happens on new coding tests and in real work. In writing and in business tasks with a clear answer, Argon is genuinely strong, even on tests that are hard to train for.

Field examples: Argon vs. Opus 5.5

On the first day, developers ran the same request on both models. Here are two checks that favor Argon, one that favors Opus, and one report about a usage quota.

A lava lamp, 18 cents against $1.27. BridgeMind runs a fixed task on every model: build a 3D lava lamp. Argon finished in under two minutes, for 18 cents, with clean glass and real-looking lava. Opus 5.5 worked for nine and a half minutes, for $1.27.

BridgeMind's post and video on X, October 13, 2026

The same request, $1.20 against $8.20. The developer J A Z I I gave both models the same request, at the highest reasoning level. In his words, the results came out "almost identical," and Argon cost 15% of the price.

notjazii's post and comparison video on X, October 13, 2026

Code and agent work: Opus is ahead. The designer jumperz wrote that he was impressed by what Argon does with visuals and complex documents, but that in code and agent work "Opus is still way ahead," and that this isn't a big enough leap.

jumperz's post on X, October 13, 2026

A week's quota in half an hour. ashen_one ran three simple tests, including a Minecraft clone, and wrote that the results were "not amazing," and that the tests burned through his entire weekly quota on a $100-a-month subscription within 30 minutes. And for anyone considering a subscription: Google has not published how much Argon usage each plan includes.

ashen_one's post on X, October 13, 2026

On Hacker News, in the launch thread, the picture is similar. Some commenters wrote that Opus 5.5 is "the consensus best coding model," and others praised Argon for clear, direct reasoning and for its speed. Some reported first-day problems: reasoning loops that never end and server errors.

What Argon really costs: a fifth, a third, or the same as Opus

The claim that Argon does the same work for a fifth of Opus's price is true in some cases. It all depends on the type of task and on the date:

Three things stand out in the chart:

• In one-off visual builds, the gap is large. In two checks Argon cost about a seventh of Opus, and in the third, by OhMyUnicorn, about 60%. Argon finishes fast and stops. Opus keeps polishing, and in OhMyUnicorn's check it reviewed the page it built three times.
• In broad work, Argon costs about a third at launch pricing, and half to two thirds at December's price. Opus was measured there at Max, the expensive level. According to Artificial Analysis, the savings come mainly from the low per-token price. Argon uses about 62K output tokens per task, half of what Opus uses, but 2.3 times as many as GPT-6 Astra.
• In complex code, the savings can disappear. On CursorBench, Argon cost $4.57 per task, against $3.97 for Opus 5.5 at a similar reasoning level. Cursor itself wrote that on this test Argon is "on par with Opus 5.5," and its score is indeed close, about 3 points lower.

And the date to mark: November 30. Until then, Argon costs $2 per million input tokens and $10 per million output tokens, the same as Claude Sonnet 5.5 and GPT-6.1 Sol. From December 1, the price rises to $4 and $20, exactly Opus 5.5's price. Anyone building a workflow on Argon today because of the price should already run the numbers at December's price.

Where Google really leads

This is not an anti-Google article. There are areas where, in my opinion, Google is especially strong, and they matter to most businesses more than coding tests do:

My opinion: why Google, with all its resources, is still behind

Here I move from facts to interpretation.

Argon is a real improvement, and I don't think Google lied. But I do think it presented Argon as a model that leads the world, and that isn't the case. Personally, when a company presents numbers that don't hold up in the field, it lowers my trust. With Google's next table, I'll read the methodology document first, and only then the table.

I believe in Google. It has the researchers, the chips, the data and billions of users. That is exactly why it's worth asking why it is still behind. My answer: Google is a public company, in a new and risky field.

• The pressure to show results. Alphabet has obligations to investors, and every launch is also judged by the stock market. According to Bloomberg, the stock rose about 1.7% in trading after the announcement and gave back part of the gain after the report on employee skepticism. When the market reacts to a table, there is a strong incentive to pick the tests and the competitors that flatter you.
• Less freedom to move fast. As a subscriber, I've seen Anthropic and OpenAI change several times how much usage each subscription gets. In June, a lawsuit seeking class-action status was filed against Anthropic, claiming that the usage limits on Max subscriptions are lower than advertised. These claims have not been decided. In my opinion, for a public company the size of Google, a move like that could cost a class action worth billions, so it is more constrained at every step. I suspect the same caution is at work in safety.
• Safety. Google opened Argon first to a closed group of cyber defenders. In a field like cybersecurity, that's the right call, but it costs time, and smaller competitors move faster.

I have one more hypothesis about the two-week delay: that the first group to work with the model saw that the results in the field were weaker than the table promised. I have no source for this. Google's official explanation is safety alone, and this is my hunch, not a fact. What is documented: Gemini 3.5 Pro, promised for June, was cancelled, and Google employees told Bloomberg that Argon is weaker at work than on tests. That report was published on announcement day, so it doesn't explain the delay itself.

Where Argon stands in Sigma AI Ranking

Argon entered Sigma AI Ranking on October 13 marked "Preliminary score," because only one source, Arena, had measured it. In the October 14 update, the price, Epoch AI's knowledge test and the full Arena rankings came in, and it received a full score. The result will probably surprise anyone who has read this far:

Argon is in second place, 0.9 points behind Opus 5.5, and under our rules that is a tie at the top. We did not change the method to make the result fit my opinion, so it's important to explain what's behind the number:

• Writing and reliability come from Arena. Argon is first there because people prefer its answers. That is its real strength, and also where the suspicion of "tuning to preference data" applies.
• Its agent score is estimated. Argon has no Terminal-Bench result yet, which is our main source for agent and coding work, so the score was calculated from parallel sources and is marked ≈ on the page. Its general capability score is also estimated, because Epoch AI has not yet published its general index for it. None of the new coding tests from the benchmaxxing section went into the calculation, because they are not openly licensed.
• The price is the launch price. The ranking calculates cost at the current price, $2 and $10. By my calculation, on December 1, when the price doubles, Argon will drop to about 88.3 and fall out of the tie.

In other words, Sigma shows exactly what this article describes: on the sources that measure writing, knowledge and price, Argon is at the top, and its gap lies where the open sources don't reach yet. When a Terminal-Bench result is published, the score will update. Why we built our own ranking, and why it relies only on open sources, we explained in the article on Sigma AI Ranking.

Who Gemini 4 Argon is for, and who it isn't for yet

Argon or Opus 5.5 for your next task?

Five questions about one task you're considering. After you see the result, run the task once on both models and check the bill.

Is the task mainly writing, summarizing or working with documents, rather than software development?

Yes No

Is a good result enough for you, even without the highest polish?

Yes No

Does the task leave out building an interface, a website or animation?

Yes No

Is the task in a language other than English?

Yes No

Does the task repeat many times, so the cost per run matters to you?

Yes No

Answer all the questions to see a result.

Four things people say about Argon, and what's true

What this means for a business owner

Three practical takeaways:

• There are now three strong models at half of Opus's per-token price: Claude Sonnet 5.5 and GPT-6.1 Sol permanently, and Argon until the end of November. For writing, documents and customer service in languages other than English, Argon is a serious candidate. Now is the time to test it.
• Don't choose a model by the maker's table. Read who measured each row and under what conditions, and test on one real task of your own, with a clock and a bill. I also wrote about this in the article on Meta's Muse Wonder: there too, the maker's table told a different story from the field tests.
• Choose a model by task, not by company. Within three months, the leading tool changed hands twice, as I described in the article on Anthropic's comeback. A good agent can work with Argon on the writing and with Opus on the code.

Platin Digital is an OpenAI Select Partner, and we build the agents we set up with the models and tools that best fit each business, from OpenAI, Anthropic, Google and others. Argon is going into our testing for writing and for work in languages other than English. For code, and for work where quality decides, I'm staying with Opus 5.5 for now.

Further reading: Gemini at Work 2026: Google's agent for business, ChatGPT, Gemini or Claude for your business and why we built Sigma AI Ranking.

Frequently asked questions

Sources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
• https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
• https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/
• https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf
• https://storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf
• https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf
• https://ai.google.dev/gemini-api/docs/pricing
• https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
• https://gemini.google/release-notes/
• https://platform.claude.com/docs/en/about-claude/pricing
• https://developers.openai.com/api/docs/pricing
• https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs
• https://artificialanalysis.ai/models/comparisons/gemini-4-argon-vs-claude-opus-5-5
• https://x.com/ArtificialAnlys/status/2105392625788637299
• https://www.vals.ai/models/google_gemini-4-argon
• https://epoch.ai/data/benchmark_data.zip
• https://cursor.com/cursorbench
• https://www.frontierswe.com/
• https://cognition.com/frontiercode
• https://x.com/arena/status/2105394855644139908
• https://simonwillison.net/2026/Oct/13/llm-gemini/
• https://news.ycombinator.com/item?id=50073879
• https://x.com/bridgemindai/status/2110050421184360718
• https://x.com/notjazii/status/2110073240404271335
• https://x.com/jumperz/status/2110064835270689125
• https://x.com/LexnLin/status/2110057390259290540
• https://x.com/ashen_one/status/2110055018191565201
• https://x.com/bindureddy/status/2110038264463987088
• https://x.com/cursor_ai/status/2110087611763368271
• https://x.com/teortaxesTex/status/2105455812915433849
• https://ohmyunicorn.com/labs/dream-loop/planet-argon/
• https://andonlabs.com/evals/blueprint-bench-2
• https://artificialanalysis.ai/evaluations/omniscience
• https://www.thestar.com.my/tech/tech-news/2026/10/02/google-grapples-with-employee-scepticism-about-new-gemini-4
• https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release
• https://www.khaleejtimes.com/business/google-announces-gemini-4-flagship-ai-model-after-months-of-delays-2
• https://www.techzine.eu/news/applications/144960/opinions-are-divided-on-gemini-4-argon/
• https://www.engadget.com/2194626/anthropic-hit-with-lawsuit-over-its-claude-max-usage-limits/