Picture two quiz contestants. One answers every question, fast and confident, and a fair share of those answers are wrong. The other passes now and then, saying “I don’t know that one,” but when they do answer, you can usually bank on it. On a quiz show the first contestant may look more impressive. In your job, writing a report your boss will forward, you want the second one.
Gemini 4 Argon is the second contestant. It is Google’s first Gemini 4 model, and the trait that sets it apart is not raw speed or flash. It is how rarely it makes things up.
What it is, and where you’ll find it
Google announced Gemini 4 Argon on September 30, 2026. At first only a small circle had it: the Fairwind Program, a group of vetted cybersecurity defenders, and then some top-tier subscribers. On October 13, 2026, Google opened it up much more widely.
Here is the part that matters for most readers. Argon is in the Gemini app and in AI Mode in Google Search for Google AI Pro (about $20 a month) and Ultra subscribers. If you use Gemini for free, or on the cheaper Google AI Plus plan, you don’t get it. Developers can call it through the paid tier of the Gemini API and AI Studio as gemini-4-argon, and choose a low, medium, or high thinking level, much like choosing how long someone may think before answering.
It replaces Gemini 3.1 Pro, which stayed a Preview model for seven months while its promised 3.5 Pro successor never arrived.
Google is back near the top
Artificial Analysis, an independent testing firm, ran its benchmarks at launch. Its Intelligence Index combines ten hard evaluations, from office work and business automation to scientific coding and factual knowledge. Gemini 4 Argon, at its high thinking level, scored 53. That ties GPT-6 Astra, OpenAI’s flagship, and sits one point above GPT-6.1 Sol. Artificial Analysis summed it up simply: Google is back among the top three labs.
To see how big a jump that is, consider that just before Argon, Google’s best-placed model on the same index was Gemini 3.8 Flash at 41.
Two other independent scoreboards agree. Vals AI, which tests models on finance, legal, tax, and coding work, put Argon first on its Vals Index at 68.9%, just ahead of Claude Sonnet 5.5 and Claude Opus 5.5 (both about 67%). On Arena’s Text leaderboard, where people compare two anonymous answers and vote for the better one, Argon, tested at its high thinking level, reached a preliminary #1 with 1525 points. Arena’s category boards show it doing especially well on hard prompts, instruction following, longer questions, and creative writing.
To be fair, Anthropic still leads Artificial Analysis’s index. Claude Opus 5.5 scored 58 and Claude Sonnet 5.5 scored 56.
The real headline: fewer made-up answers
Every chatbot has the same weakness. It doesn’t look facts up by default; it predicts the most likely next words. When it doesn’t know something, the most likely-sounding answer is often a confident invention.
Artificial Analysis measures this with a test called AA-Omniscience. It asks hard factual questions and checks what the model does when it doesn’t know. Of the questions Gemini 4 Argon did not answer correctly, only 15% got a confident wrong answer; the rest got an honest “I don’t know.” That is the lowest rate of any model scoring 45 or more on the Intelligence Index. For comparison, GPT-6 Astra guessed wrong in 51% of those cases, GPT-6.1 Sol in 54%, and Claude Sonnet 5.5 in 47%. Note what this number is not: it is not the share of all its answers that are wrong.
In practice, this means Argon will say “I’m not sure” more often. There is a trade-off. On the same test, Argon answered 50% of all questions correctly, against 63% for GPT-6 Astra, and its overall Omniscience score (42) was a point below Astra’s (43). Argon doesn’t know more; it guesses less. For knowledge work, where a confident mistake can travel far, that is often the better trade, but keep a second source handy for the questions it declines.
What it’s good at in a working day
- Long documents and research. Argon has a one-million-token working space, enough for several thick reports at once. That space is shared between what you give it, its thinking, and its answer, so a very long question leaves less room for the reply.
- Mixed media. It reads text, images, audio, and video, and answers in text. Google reports 91.7% on LVBench, a test of understanding long videos. A researcher can hand it a recorded interview, a chart, and a report and ask what connects them. To create pictures, the Gemini app uses Google’s separate image models.
- Multi-step office chores. On Artificial Analysis’s run of AutomationBench, which tests end-to-end business workflows across many tools, Argon placed first at 77.5%.
- Writing that sounds less like a robot, according to Google. Google says Argon hedges less, repeats itself less, and uses fewer clichés than older Gemini models, and follows a requested style more consistently. That is Google’s claim; we have not yet seen an independent test of it.
The honest catch
No free version. For a model aimed at everyday work, this is the biggest limit. Google’s free and Plus tiers don’t include it, and neither does the free API tier.
The low price won’t last. The API costs $2 per million input tokens and $10 per million output tokens until November 30, 2026. After that, it doubles to $4 / $20. Artificial Analysis measured about $1.99 per task at the launch price, around 60% of Astra’s cost, but that becomes about $4 at the standard price, well above GPT-6.1 Sol’s $0.72.
It talks a lot. Artificial Analysis needed 110 million tokens to run its whole index with Argon, against a median of about 85 million, and called the model “somewhat verbose.” The thinking it does before answering is billed as output, so long answers cost more and take longer than the price per token suggests.
Polish can lag behind the thinking. A hands-on review by Stork.ai found Argon’s design work looked cramped next to rivals and that it worked best with a human steering it. Expect strong analysis, and plan one more pass for a client-ready deliverable.
Some numbers are Google’s. The long-video score and Google’s comparison table, where Argon tops 13 of 19 rows against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5, come from Google. The independent results from Artificial Analysis, Vals AI, and Arena are the more reliable guide, and they are good.
Who should use it
| If your day looks like… | Reach for | Why |
|---|---|---|
| Research and long documents where errors are costly | Gemini 4 Argon | Least likely of the leading models to guess on an independent knowledge test |
| You already pay for Google AI Pro and live in Google’s apps | Gemini 4 Argon | It’s in the Gemini app and in AI Mode in Search |
| The highest independent scores, regardless of cost | Claude Opus 5.5 | Highest on the broad independent index |
| Strong everyday help without a subscription | Claude Sonnet 5.5 | On Claude’s free plan |
| Cheap, high-volume document work in ChatGPT | GPT-6.1 Sol | Near-flagship results at a much lower cost per task |
The everyday verdict
Gemini 4 Argon is Google’s comeback, and it comes back with a useful personality: smart enough to tie OpenAI’s flagship on an independent index, and humble enough to admit when it doesn’t know. If you already pay for Google AI Pro, make it your default for research, long reports, and anything where a made-up fact would embarrass you. If you build on the API, enjoy the launch price, but plan your budget for the doubled price that starts in December.
