Can Pangram detect Gemini 4 Argon?

Gemini 4 Argon is Google’s latest AI model, released October 13, 2026. Google claims that Gemini 4 Argon was built for complex, multistep tasks rather than chat, and that it writes in a straightforward manner and hedges less often than its predecessors.

With these writing style changes, does Pangram still detect it? We tested Pangram 4.1’s performance on 2,341 example prompts and found that Pangram 4.1 classified 16 Gemini 4 Argon outputs as human, 7 as mixed, and 2,318 as fully AI-generated, achieving an accuracy of 99.02% on Gemini 4 Argon.

To generate the benchmark set, we used 1,200 first-turn prompts from the LMSYS-Chat-1M dataset (Zheng et al., 2023). Then we gave those to Gemini 4 Argon on medium and high reasoning and ran the output through Pangram 4.1.

Results by category

Overall, Gemini 4 Argon had a false negative rate of 1.28% out of 1,174 responses on high reasoning, and a false negative rate of 0.69% out of 1,167 responses on medium reasoning. Overall, there was not a significant difference in detection rate between medium and high reasoning (p=0.21).

High reasoning

When Gemini 4 Argon was set to high reasoning, Pangram 4.1 categorized 11 responses as human, 4 as mixed, and 1,159 as fully AI, achieving a recall rate of 98.72%.

Medium reasoning

On medium reasoning, Pangram 4.1 categorized 5 responses as human, 3 as mixed, and 1,159 as fully AI, achieving a recall rate of 99.31%.

How does Gemini 4 Argon write?

Google asserts that Gemini 4 Argon hedges less and speaks more directly. For a rough test of this, we look at a set of 101 hedge words sourced from Hyland (2005), and compared the occurrence of those words to the occurrence in Haiku (7.0 per 1,000 words) and GPT-6.1 Sol (7.5 per 1,000 words).

Below is the list of hedge words we used:

about, almost, apparent, apparently, appear, appeared, appears, approximately, argue, argued, argues, around, assume, assumed, broadly, certain amount, certain extent, certain level, claim, claimed, claims, could, couldn't, doubt, doubtful, essentially, estimate, estimated, fairly, feel, feels, felt, frequently, from my perspective, from our perspective, from this perspective, generally, guess, indicate, indicated, indicates, in general, in most cases, in most instances, in my opinion, in my view, in this view, in our opinion, in our view, largely, likely, mainly, may, maybe, might, mostly, often, on the whole, ought, perhaps, plausible, plausibly, possible, possibly, postulate, postulated, postulates, presumable, presumably, probable, probably, quite, rather x, relatively, roughly, seems, should, sometimes, somewhat, suggest, suggested, suggests, suppose, supposed, supposes, suspect, suspects, tend to, tended to, tends to, to my knowledge, typical, typically, uncertain, uncertainly, unclear, unclearly, unlikely, usually, would, wouldn't

Gemini 4 Argon used hedges 51% less than Haiku and 54% less than Sol when measured per 1,000 words. Funnily enough, even though models in general tend to use hedge words much more frequently than humans do, this list of words is detected as fully human by Pangram 4.1. For fun, I also asked ChatGPT to rearrange this list in any order it would like, and that list was also detected as fully human.

Gemini 4 Argon had the longest answers of those three models, with a median of 447 words (medium reasoning 418 words, high reasoning 447 words) compared to 371 for Haiku and 313 for Sol. Gemini also refused to answer 31 prompts across all 2,400 requests, plus 3 requests that were blocked by Google’s filter, compared to 35 refusals for Haiku and 23 for Sol.

Earlier benchmarks

Pangram is robust to Gemini 4 Argon on natural language prompts, and detects it with an observed accuracy of 98.72% on high reasoning. It detects Gemini 4 Argon on medium reasoning with an observed accuracy of 99.31%, for a combined 2,318 out of 2,341 (99.02%) across both settings. Pangram is able to generalize to new models because new model releases tend to inherit much of their predecessors' style and voice, so Pangram doesn't need to be retrained for every new release.

Here's how Gemini 4 Argon compares to our benchmarks of other recent frontier models:

If you want to check a specific document, you can try Pangram here.