Google's new AI guessed wrong on 15% of questions it couldn't answer. OpenAI's 51%.
Google launched Gemini 4 Argon on Wednesday.
Independent testing firm Artificial Analysis scored it level with OpenAI's GPT-6 Astra on its intelligence index, both at 53.
The gap appears when a question is beyond the model. On the firm's hallucination test, Argon guessed wrong instead of declining on 15% of the questions it could not answer correctly. GPT-6 Astra did so on 51%, and OpenAI's GPT-6.1 Sol on 54%.
"Gemini 4 Argon is much more likely to acknowledge when it does not know an answer rather than guess incorrectly," Artificial Analysis wrote.
Google is charging an introductory $2 per million tokens of input and $10 per million of output, half its list price. Tokens are the chunks of text AI models are billed by.
Access is limited for now. Selected cybersecurity teams get it first. Google says paid API customers and Google AI Ultra subscribers come next.
For the work your team hands to AI, what is a model that says "I don't know" worth to you?
Image: AI-generated.
Sources
Our file on Google
- 7 Oct
Google signed for 890 megawatts from 11 old nuclear plants. Nobody builds new ones.
- 5 Oct
Google paused its open-source bug bounty after an AI report flood. Most were invalid.
- 1 Oct
Google pays about 100 publishers when their pages shape AI answers.
- 1 Oct
Trump ordered federal agencies to rename AI. The new word is Super Intelligence.
- 30 Sept
Google challenged two EU orders that help rival AI in court.
Every story here is open to read. The ERP LEADERS brief goes one step further.
The week in enterprise software, in one email: the stories that mattered, the failures with their figures, one case taken apart. Read a sample or sign up for the brief.
Welcome back. · Issue 01
