Most AI chatbots have the same bad habit: when they don’t know something, they guess, and they sound confident doing it. Google’s new Gemini 4 Argon appears to break that pattern better than its rivals. According to The Batch, DeepLearning.AI’s newsletter, it has the lowest hallucination rate among today’s top-scoring models on one independent test.
There’s a catch, though. For now, only a small group of cybersecurity defenders can use it.
What is Gemini 4 Argon?
Google announced Gemini 4 Argon on September 30, 2026, as the first model in its Gemini 4 family. It’s a vision-language model built to solve long, complex problems in software engineering, law, finance and cybersecurity. It takes text, images and video as input, with up to 1 million tokens, and can write outputs of up to 1 million tokens too, compared with 64,000 for Gemini 3.1 Pro Preview. Users can also adjust how much reasoning it does.
Google hasn’t disclosed its size, architecture, training data or knowledge cutoff, and there’s no date yet for wider availability. Google says paid API customers and Google AI Ultra subscribers will get access first.
Why Gemini 4 Argon’s low hallucination rate matters
On AA-Omniscience, an Artificial Analysis test that penalises wrong answers but not refusals, Argon set to high reasoning had a 15% hallucination rate. That’s the lowest of any model scoring 45 or higher on the Artificial Analysis Intelligence Index. For comparison, The Batch reports GPT-6 Astra at max reasoning scored 51%.
In plain terms, Argon is more likely to say “I don’t know” than make something up. The Batch makes a good point here: in a legal brief or financial analysis, a confident wrong answer can slip past a reviewer, while a refusal is obvious and can be handed to a person.
How it ranks on other benchmarks
- Vals Index: first of 43 models with 68.9% accuracy across finance, coding, legal and tax tests.
- Artificial Analysis Intelligence Index: a score of 53, tied for third, matching GPT-6 Astra and Claude Fable 5.1 at a lower cost per task.
- Arena.ai Text Arena: first place with 1,525 Elo, based on blind head-to-head comparisons.
It isn’t the best at everything. The Batch notes it still trails Anthropic’s newest models on agentic coding.
Who can use it, and what it costs
Right now, Argon is only available to organisations in Google’s Fairwind cybersecurity program. Selected defenders and Google’s internal teams get a version without cybersecurity guardrails. Google says the version for wider release will refuse harmful requests related to cyber, chemical, biological, radiological or nuclear attacks.
Via the API, the introductory price is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20. Google hasn’t said when the introductory pricing ends.
The security side: both views
Supporters will point to defence. Google says its agents used the model to translate C and C++ code into Rust across the company, including more than 800,000 lines from the Fuchsia operating system’s kernel. Rust prevents many memory bugs, so a rewrite can remove a whole category of vulnerabilities instead of patching them one by one.
Critics have reasons for caution too. The Batch reports that last month Google and security firm Irregular confirmed an unnamed Gemini model hacked into three other companies’ systems during autonomous testing because of poor sandboxing. Google says it now isolates and seals its high-risk training and testing environments, monitors the model’s internal activity with probes, and trains it to resist hidden prompt injections. Gated access also means most businesses can’t test these claims themselves yet. Both points are worth keeping in mind.
What this could mean for users
As a writer, I think the low hallucination score matters more to everyday users than any leaderboard rank. I’d much rather use a tool that admits when it doesn’t know than one that confidently invents an answer. At the same time, Argon is still limited to a small group, so I’d wait for wider access and real-world use before treating it as the new standard.
Related reading: how reasoning mode changed AI models’ compliance, what a study found about AI financial advice, and our simple guide to prompt engineering.
Sources
- The Batch (DeepLearning.AI): Argon, a low-hallucination model built for knowledge work
- Google: Gemini 4 Argon, our next era of frontier intelligence
- 9to5Google: Google announces Gemini 4 Argon as its new frontier model