On 13 October Google opened Gemini 4 Argon to everyone. The press release calls it the company's most powerful model, and the headlines say Google has overtaken the field. I went to the numbers, because at NeuPortal we work with models on real markets every day and stopped believing "most powerful" without a table next to it a long time ago.
In short: Argon really is in first place. But that first place looks nothing like the way it is being sold.
What launched, and when
It rolled out in two stages. On 30 September only Google's cybersecurity partners got it, through the Fairwind Program. Google says plainly that Argon was trained partly for defensive security work: finding, validating and patching vulnerabilities in code on its own. On 13 October access opened to everyone: developers through the API and AI Studio, Pro and Ultra subscribers in the Gemini app and in AI Mode in Search.
The context window is one million tokens, input and output combined, with answers of up to 262,000 tokens. That is a lot: a large codebase or a stack of several years of reports fits in a single request.
First place - by how much
Google points to the independent Vals AI index. It tests models not on school puzzles but on real work: finance, law, tax, code. And there Argon really is first, at 68.9%.
Now look at who is next to it:
- Gemini 4 Argon - 68.9% - Claude Sonnet 5.5 - 67.0% - Claude Opus 5.5 - 67.0% - Claude Fable 5.1 - 65.8%
The gap between first and second is 1.9 points. The margin of error on Argon's own score is about one point. So "Google has overtaken everyone" is more honestly put as "Google has joined a top three that is running almost level". That takes nothing away from the model. It tells you where the race is: the jumps of tens of points are over, and the lead now changes hands with every release.
Where Argon is genuinely strong
Look past the overall index at the individual Vals tests and the picture gets more interesting:
- **Finance (Finance Agent):** 65.4% - first of 78 models. For us this is the line that matters most: analyst work, reports, calculations. - **Competitive programming (IOI):** 100% - shares first place with GPT-6 Astra. - **Building apps from a description (Vibe Code Bench):** 91.9% - second, just behind Claude Sonnet 5.5. - **Code migration:** 68.2% - also second. - **Cybersecurity (CyberBench):** 77.9% - third. A curious detail: the model launched as a "cyber model" is not first here; GPT-6 Sol is ahead. - **Legal tasks (LegalBench):** 88.3% - third of 151.
Google adds figures of its own that nobody has independently checked yet. In its Antigravity development environment, accepted code changes rose by more than 35% and rollbacks fell by nearly a third. The model sped up the xTB quantum chemistry package by up to 2.3 times, and taught the Apollo humanoid robot to play cornhole from a single video - seven hits in a row in a real-world trial. Google also highlights its defence against prompt injection: by its own figures, only 0.7% of attacks succeeded.
Where Argon is weaker
The press releases skip this, but the index shows it:
- **Using a computer (CUA-bench):** 4.8% - eighth of nine. A model that writes code brilliantly is still poor at clicking through interfaces in place of a person. - **Medical notes (MedScribe):** 17th of 109. - **Large engineering tasks (ProgramBench):** 2.5%.
And the most practical point: one task in the Vals index costs $15.68 on average, and an answer takes 46 minutes on average. Argon thinks long and expensively. For heavy analysis that is fine; for quick answers it is not.
What it costs
Until 30 November there is an introductory price: $2 per million input tokens and $10 per million output tokens. From 30 November it is $4 and $20. Cached input tokens are 95% cheaper. If you plan to build a product on Argon, do the maths at the full price from the start: the introductory one ends in six weeks.
What it means if you work with markets
At NeuPortal we do not pick "one best model". We test models on our own tasks and publish the results, misses included. Argon interests us mainly for its first place on the finance test and its long context: it is a strong model for working through filings, research and large datasets.
But I would keep two conclusions in mind:
1. **Leadership is now a matter of a couple of points.** The model that does best on a general test set is not necessarily the best on your task. Test it on yours. 2. **A strong model is not a trading signal.** Even the best model makes mistakes, and in a market a mistake costs money. That is exactly why we keep a public score of our forecasts against Polymarket - and the market is still often more accurate than we are: neuportal.ai/experiment
The bottom line
Gemini 4 Argon is a strong release and an honest first place on an independent index. It is especially good at finance, code and tasks that need long, deep reasoning. But the gap to Claude and GPT-6 is a handful of points, and it trails on working with interfaces and on response speed. Google is back in the lead. Until the next release from its rivals.
Follow our breakdowns and our public forecast score in the Telegram community: t.me/neuportal_community
*Not financial advice. Sources: Google's blog, TechCrunch, Vals AI (data as of 14 October 2026).*