Gemini 4 Argon vs Opus 5.5: I Gave Both the Same Repo

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Gemini 4 Argon vs Opus 5.5: I Gave Both the Same Repo · Channel: Prompt Engineer 48 · 08:20 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI models, Open source and software ecosystems

Summary

The video compares Gemini 4 Argon and Claude Opus 5.5 on five coding tasks using the same Python tabulate repository, prompt, and one-shot setup. Both models fixed the first two bugs and handled a more complex task involving three bugs, while Argon’s patch for a boolean-column bug had malformed formatting. On the feature task, Argon failed hidden tests and an existing signature test, whereas Opus passed all 392 tests. The creator concludes that Argon kept up on bug fixing in this small test, but Opus handled the feature specification and edge cases better.

Key points

Google describes Argon as a frontier model for coding, enterprise knowledge work, and cyber defense, with text, image, audio, and video inputs.
The video notes that Google’s reported benchmark scores and internal-use examples could not be independently verified by the creator.
The Deep Search leaderboard cited in the video was last updated before Argon was announced, so Argon was not listed there.
The test used the open-source Python tabulate repository, with the same files and prompt provided to both models and no second attempts or tools.
Both models fixed the off-by-one decimal-alignment bug and the missing-value-text bug.
For the boolean-column bug, Argon made the intended one-word change but produced a malformed diff that required fuzz to apply.
Both models fixed three planted bugs across different functions and passed the existing 382-test suite.
On the max_rows feature, Argon failed three tests, including two hidden-test crashes and an existing argument-count test; Opus passed all 392 tests.
The creator reports Opus cost $2.08 across all five tasks; the Gemini app did not show enough information to calculate Argon’s cost.
The creator cautions that the sample was small and says Argon might perform differently in a workflow where it can run tests and iterate.

Chapters

Start	Chapter	Summary
00:00	Introducing Gemini 4 Argon	The creator outlines Argon’s reported capabilities, benchmarks, rollout, and limitations before setting up a coding comparison.
03:09	The repository test setup	The creator explains the pricing context and describes the one-shot comparison using the Python tabulate repository.
04:11	Bug-fixing tasks	Both models pass the first two bug tasks, Argon produces a malformed diff on another, and both handle three simultaneous bugs.
05:44	The max_rows feature task	Argon fails hidden and existing tests due to a variable-order logic bug and an unupdated argument-count test, while Opus passes the full suite.
07:17	Results and caveats	The creator summarizes the scores, notes the test’s limitations, and recommends checking model-generated feature changes against tests.

Quotes

Gemini 4 Argon got four of them right. Claude Opus 5.5 got all five. — 00:00

Argon made the right one-word change, but its diff was malformed. — 04:43

Python reads the condition first, so the variable does not exist yet, and you get an unbound local error. — 06:14

For a spec with edge cases and keeping an existing suite green, Opus was better in my run. — 07:48

Concepts and entities

Gemini 4 Argon, Claude Opus 5.5, Python tabulate, Claude Code, Gemini app, Antigravity, RunPod, Google, Anthropic, Deep Subly

Full transcript (original language)

00:00 Five coding tasks, same repo, same prompt for both models. Gemini 4 Argon got four of them right. Claude Opus 5.5 got all five. And the one Argon missed is the one that matters most. Let’s go ahead and see exactly what happened. So, what actually is Gemini 4 Argon? Google calls it its frontier model for real-world coding, enterprise knowledge work, and cyber defense. The announcement post is dated September 30th, and the model card was published on October 13th. According to the model card, it is a sparse mixture of experts

00:31 transformer. It takes text, images, audio, and video as input. It outputs text only, and its knowledge cut-off is March 2026. Now, the size Google’s blog says the output token limit goes up to 1 million tokens, up from 64,000. So, it is built for very long runs, where the model keeps working on one problem for a long time. The model card describes it a bit differently, a combined input and output window of up to 1 million two descriptions. So, check the docs before you plan around it. What does Google say it already did with

01:02 Argon inside the company? Three examples are on this page in quantum computing. It beat a published baseline by 40% in a matter of minutes. Its agents found memory optimizations across the data centers that free up over 300 TB. And for a video decoder called libgav1, it wrote Rust that runs 2.7 times faster than the existing Rust port with identical output. These are Google’s own reports on its own systems, and I cannot check them. Now, the benchmarks on Deep Subly version 1.1, it scores 77.9%.

01:34 This is Google’s own chart. GPT-6 Astra is at 74.1. Claude Opus 5.5 is at 74.2. And Claude Fable 5.1 is at 67.4. So, on Google’s chart, Argon is 3.7 points ahead of Opus 5.5. Google also reports first place on Zapier’s automation bench with 51.3% and 91.7 on LV bench, which is long video understanding. Again, all of this is Google’s own reporting. Then there is the cyber side. Argon is

02:06 going first to trusted cyber defenders through what Google calls the Fairwind program. And Google says it will remove the cyber guardrails for those defenders. The model card is also direct about it. It says Google cannot rule out that Argon reaches its critical capability levels for cybersecurity and for chemical, biological, and nuclear risk. That is the reason for the careful rollout. And for us, there’s the app. On October 13th, Google AI Pro and Ultra subscribers got Argon in the Gemini app. Google says the answers are more direct

02:38 with less hedging. There’s a toggle from chat to tasks where Gemini keeps working when you are away from your computer. Anti-gravity’s coding interface is now inside the Gemini macOS app, and the model card lists the usual limitation. It can hallucinate. So, that is Argon in general. Now, the claim I want to test Now, this is the actual Deep Search leaderboard, not Google’s chart. It has 113 tasks across 91 repositories, and it was last updated on September 22nd, which is 8 days before Argon was

03:09 announced. So, Argon is not on this board yet. The best models on it sit in the low 70s, and that 77.9 is a number I cannot confirm myself. Now, the price on the Gemini API pricing page, Argon is $2 per million input tokens and $10 per million output tokens through November 30th. Then from December 1st, it is 4 and 20, so it doubles. Cached input is 10 cents per million now, 20 cents after. And see this note at the top. It says Argon is

03:39 not available on the free tier or on the Google AI Pro plan for the API. So, let’s test it. I took Python tablet, a real open source repo, MIT license, about 2,600 stars with 382 passing tests. Then I broke it in four different ways and gave it one new feature to build and the repo’s own tests decide who passes. Same prompt, same file, same pretest output to both models. Opus 5.5 gets it through Claude code with tools turned off, so it is one-shot just like Argon in the Gemini app. Each answer has

04:11 to be a diff and I apply that diff to a fresh copy and run the whole test suite. No second tries, no hints, and the files I attach are the same for both. Task A is an off-by-one bug in how decimals are aligned. 97 tests fail. Both models found it and fixed it one line and all 382 tests pass. Task B, missing value text gets ignored. Five tests fail. Both fixed it, both pass. So far, a tie. Task C, a column of booleans gets

04:43 treated as integers. Four tests fail. Opus fixed it cleanly. Argon made the right one-word change, but its diff was malformed. So Git refuses to apply it. It only applies with fuzz. Now, I also have to be honest here. The first time I ran this, the attachment did not reach Gemini. That was my script and Argon guessed. I ran it. So this is the second attempt. Task D is where I made it harder. Three bugs in three different functions at once. 103 failing tests. Both models found all three. Argon replied in about

05:13 30 seconds. Opus took about 23. Both passed 382 of 382. And at this point, honestly, Argon looks very good. But there is one more task and this is the one where they split. Hold on to this one. Task E is not a bug. It is a feature. I asked both models to add a new argument to tabulate called max_rows. Keep the first n rows and add a footer line saying how many rows are hidden.

05:44 Separating lines after the cutoff have to be dropped and the existing test suite must still pass. Here is Argon. I attach the two files, paste the prompt, and send it. It generates for 28 seconds, and there is the diff. Now, here’s how I score it. I run 10 hidden tests that I wrote and never showed either model, plus the full existing suite. Argon fails three. Two of my hidden tests crash with an unbound local error, and the existing signature test fails

06:14 two. Opus all 10 tests and the full suite 392 tests green. So, why did Argon crash? Look at this list comprehension that is supposed to filter the separating lines. It uses the variable I in the condition, and I only gets defined in the for loop that comes after it. Python reads the condition first, so the variable does not exist yet, and you get an unbound local error. It is a logic bug that only shows up when a table has a separating line. And the second miss is simpler. Tabulate has

06:46 a test that counts its arguments. 20 Argon added a 21st argument and left that test at 20, even though my prompt said to update it if needed. Opus updated it. Cost for Opus, I have real numbers from the CLI between 38 and 42 cents for the first four tasks, and 50 cents for the feature task. $2.08 for all five. For Argon, the Gemini app shows no token count, so I cannot give you a real cost, and I’m not going to make one up.

07:17 A few limits you should know about. Each model got one run per task, and these are bugs I planted, not real issues from the repo history, and both models answered one shot with no tools. Argon inside anti-gravity, where it can run the tests and loop on the failures, might well fix that feature task. I did not test that, so I’m not claiming it. So, the scoreboard. Opus 5.5, five out of five. Argon four out of five with one patch that needed fuzz. Four broken variants and one feature is a small sample and it says nothing about 77.9 on

07:48 Deep Swell. What it does say is this, for finding bugs Argon kept up even with three at once. For a spec with edge cases and keeping an existing suite green, Opus was better in my run. If you mostly fix bugs and you have pro, try Argon in the app. It is real. If you hand a model a feature spec, check its diff against your tests. If you’re looking for cheap and reliable GPUs, I use RunPod and the link is in the description. So, in this video we put Gemini 4 Argon against Claude Opus 5.5 on Python tabulate. Tell me in the

08:20 comments which kind of task I should break next. I’ll see you in another one. Check out the next video.