18 points. That is the gap one developer found when Claude Opus 5.5 and Gemini 4 Argon each wrote a plan for the same real feature. Claude's plan scored 77 out of 100. Argon's scored 59. A second Reddit user ran 12 hard tasks, from tricky bugs to real security patches, and saw a very different picture: Argon tied Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol on 11 of them, at roughly a third of Sonnet's cost.
Both are single runs by anonymous users. Neither is a benchmark. But set next to Google's own scorecard and an API price that is scheduled to double on November 30, they show how much room sits between a launch table and a working day.
This article is built from a YouTube breakdown of the Argon rollout, checked against Google's announcement and independent trackers. Where I rely on a secondary report, I say so. For background on what Google had confirmed before launch, see my earlier look at Gemini 4. For the other side of this matchup, there is the Opus 5.5 leak that was retracted.
Where Argon Sits on the Independent Indexes
Artificial Analysis runs 10 tests across math, science, coding and reasoning on its own hardware, so no lab grades its own homework. On that index Argon scored 53. That ties GPT-6 Astra, sits 1 point above Meta's Muse Wonder, and lands about 5 points behind both Claude Opus 5.5 and Sonnet 5.5. Google's previous best model scored around 30, which makes the jump roughly 23 points in one generation. That last number is the video's arithmetic on an approximate figure, so read it as a ballpark. One tracker also notes the Argon result is for its high reasoning setting.
A second index tells a different story. On the Vals Index, a GDP-weighted ranking of professional tasks, one comparison write-up reports Argon first at 68.9%, Sonnet 5.5 at 67.04% and Opus 5.5 at 66.97%. That is a lead of about 2 points, built on a different mix of work.
| Index | Gemini 4 Argon | Claude Opus 5.5 |
|---|---|---|
| Artificial Analysis Intelligence Index | 52.6 | 57.6 |
| Vals Index | 68.90% | 66.97% |
The two indexes do not contradict each other so much as measure different things. The video lists where Argon leads: first place on a test of real finance, legal and tax work, and a margin of about 9 points over every rival on a Zapier test where AI finishes whole business tasks from start to end. Coding is split. Argon set a record of about 78% on Google's main test for long software projects, then lost the command-line test to all three rivals, with Claude on top.
Google's Scorecard: 18 Tests, a Tally That Adds to 20
Google's own comparison covers 18 tests and says Argon wins 14, ties 1 and loses 5. Add those and you get 20.
Other outlets describe the table differently. One headline puts Argon ahead in 12 of 18 tests against GPT-6 Astra and Opus 5.5. Another tracker counts 14 wins across 19 rows, one of them a tie. Those may reflect different comparison sets. I could not open Google's post to check the table myself, so I can't say which count is right. A reader cannot rebuild the headline from the headline, and that is the problem.
A second issue came from Mashable, as reported in the video: the scores Google lists for rival models do not match what those rivals published. In Claude's case, Google listed a higher score than Anthropic did. The video adds that Anthropic itself has said small benchmark gaps now tell you little about real-world use.
One safety number is worth a flag. The video cites a test where Argon fell for hidden instructions on websites or in emails 0.7% of the time, against 1% for Opus 5.5 and 8.5% for GPT-6 Astra. It does not say who ran the test. Treat it as Google-reported until you check the model card.
The Price Gap Has an Expiry Date
Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input 95% off. A footnote in Google's announcement post says the price becomes $4 and $20 once the introductory period ends. Coverage of the October 13 rollout puts that end date at November 30, 2026, and The New Stack notes the new rate matches Opus 5.5 per token.
| Cost measure | Gemini 4 Argon | Claude Opus 5.5 |
|---|---|---|
| API price per 1M tokens (input / output) | $2 / $10 through Nov 30, then $4 / $20 | $4 / $20 |
| Artificial Analysis cost per index task | $1.99 (about $3.98 after the price change, one outlet's estimate) | $5.98 |
Per task, the gap is about 3x today ($1.99 against $5.98). Argon used about 62K output tokens per index task in that run, and thinking tokens are billed as output. If Argon's token use holds, it stays below Opus after the doubling. If a harder prompt makes it think longer, the bill moves with it.
Access is narrower than "everyone" suggests. Google released Argon on September 30 to the U.S. government and vetted cyber defenders through its Fairwind Program. The wider rollout began October 13: the Gemini app for AI Pro and Ultra subscribers, the paid tier of the Gemini API, and AI Studio and Antigravity for Ultra users, according to coverage of Google's release notes. Free users are not on that list. Google's model page lists a context window of 1,048,576 tokens.
Two Reddit Tests, Opposite Results
The first test came from a developer on Google's top Ultra plan. They took a real feature for their live app and asked Opus 5.5 and Argon to each write a plan. GPT-6 Astra then graded both out of 100. Claude got 77, Argon 59. When shown the scores, Argon agreed with the critique. Its plan had basic design mistakes, and it fixed them once they were pointed out. The developer's verdict was that Argon needs a higher thinking setting to compete. The whole task used 2% of their 5-hour usage limit.
The second test came from a different user. They ran 12 hard challenges, from fixing tricky bugs to patching real security holes, against Claude Sonnet 5.5 and GPT-6.1 Sol. Argon tied them on 11. It was also the fastest at around 2.5 minutes per task, and the whole run cost about $3 against about $10 for Sonnet.
The one loss is the most informative result in either thread. On a long counting puzzle, Argon's output budget is about half of what its rivals get. At maximum thinking it spent the entire budget on reasoning and returned a blank answer twice. On a lower setting it gave a confident wrong number, wrapped in a fancy-sounding method. The other two models got it right.
Commenters raised the obvious objections. One called AI grading AI a "LLM centipede." Another pointed out that these models give different answers on every run, so one test proves little. The summary many landed on: not Opus 5.5 level, but fast, cheap and good enough for most work. The announcement did not clarify who these posters are, and the video does not link the threads, so I have not been able to check either one.
What Google Says Argon Did Inside Google
Google says teams using Argon ship over 35% more code, and changes that had to be rolled back fell by almost a third. It says the Chrome team used Argon to untangle about 6 million lines of old browser code, finishing months of work in weeks. Those are Google's numbers about Google's own teams.
The detail I trust most is the less flattering one, reported by Interesting Engineering and relayed in the video. Google's team gave Argon one paragraph of instructions, roughly "make our software faster without breaking anything," and left. Argon worked alone for 11 hours, went through about 205,000 lines of code, and returned 6 ideas. Humans accepted 2. Those two made long sessions about 40% faster to set up and the software 21% faster to start.
Two of six is a real result and a modest one. Early business testers quoted in the launch coverage were warmer, including a finance company that said Argon matched human analysts' numbers more than 99% of the time. Those are company testimonials, not independent measurements.
My Take
Three numbers. 5 points: where Opus 5.5 leads Argon on Artificial Analysis. 3x: Argon's per-task cost advantage today. 2 of 6: ideas humans accepted from the 11-hour run.
Argon is competitive. It is not dominant. It leads on work-style tests like finance and legal tasks, trails on general reasoning, and its price edge narrows on December 1. The Reddit tests disagree with each other, and both turned on the thinking setting and the output budget.
Nobody should pick a model from a vendor's table.
- Artificial Analysis puts Argon about 5 points behind Opus 5.5. Vals puts it about 2 points ahead.
- The $2/$10 introductory price is announced to end November 30, after which it is $4/$20, the same per-token rate as Opus 5.5.
- Google's "14 wins, 1 tie, 5 losses" on 18 tests does not add up, and outlets report different counts.
- Two Reddit tests, a plan scored 77 to 59 for Claude and a tie on 11 of 12 tasks, are single runs and show the thinking setting matters.
FAQ
Is Gemini 4 Argon better than Claude Opus 5.5?
It depends on the measure. Artificial Analysis has Opus 5.5 ahead by about 5 points. Vals has Argon ahead by about 2. Argon leads on finance, legal, tax and business-task tests, while Claude leads the command-line test. The two Reddit tests split one each way.
How much does Gemini 4 Argon cost after November 30?
Google's footnote lists $4 per million input tokens and $20 per million output tokens after the introductory period. Coverage of Google's rollout post gives November 30, 2026 as the end date. Check Google's pricing page before you commit to a budget.
Can free Gemini users use Gemini 4 Argon?
Coverage of Google's October 13 rollout lists AI Pro and Ultra in the Gemini app and the paid Gemini API tier. Free access is not mentioned. Confirm against Google's release notes, since availability can change by plan.
Is Gemini 4 Argon cheaper than Claude Opus 5.5 per task?
Today, yes: Artificial Analysis measured $1.99 against $5.98 per index task. After the price change, one outlet estimates about $3.98 per task for Argon, still below Opus. That holds only if Argon's token use stays near 62K output tokens per task.
Conclusion
Everything above is a snapshot from mid-October 2026. The price changes on November 30, index scores will move as more settings are tested, and I could not open Google's benchmark table to confirm its count. The two Reddit tests are single runs, and one was graded by another AI. Google also says the rest of the Gemini 4 family is still coming and gave no timeline.
If you are choosing between Argon and Opus 5.5 this month, the test that settles it is your own task, run on both, with cost logged next to quality.
0 Comments