Gemini 4 Argon benchmark results: accuracy vs. abstention
Artificial Analysis ● Covered by 18 sources
Google’s Gemini 4 Argon matches GPT-6 Astra on a key benchmark and is cheaper per task right now. But the real twist is it also hallucinates less, even while giving up some accuracy.
Based on reporting by Artificial Analysis — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google’s Gemini 4 Argon is back in the serious-model game. On Artificial Analysis’ Intelligence Index, the model hits 53 with high reasoning, tying GPT-6 Astra and edging GPT-6.1 Sol by one point. For Google, that puts a non-Flash model back near the top after more than seven months without one above the Flash class.
The score is only half the story. Google is leaning hard on price. With a temporary 50% discount and a 95% cache discount, Gemini 4 Argon lands at $1.99 per Intelligence Index task. That is 60% of GPT-6 Astra’s cost per task, even though Argon actually uses more output tokens on average: 62k versus 27k for GPT-6 Astra. Once the launch promotion ends, the figure rises to $3.98.
The model’s best showing may be in agentic work, where Gemini has often looked weaker than rivals. Argon ranks first on AutomationBench-AA at 78%, seven points ahead of Claude Sonnet 5.5. On Terminal Bench 4 it scores 57%, far above Google’s earlier Gemini 3.1 Pro Preview and behind Claude Sonnet 5.5, Claude Opus 5.5, and GPT-6 Astra. On AA-Briefcase, it reaches 1494 Elo, helped by a 65% rubric pass rate, though its analytical and presentation scores trail that strength.
Then there’s the weirdly important part: Argon is the most cautious of the top models in this slice. On AA-Omniscience, it posts a 15% hallucination rate, the lowest among models scoring 45 or higher on the Intelligence Index. Its raw accuracy is lower, at 50%, and that is below GPT-6 Astra. But the overall Omniscience score still comes out level with the best of the group, which says a lot about where the gains are coming from.
Google is also shipping a big practical package here: a 1M-token context window, text/image/video/speech input, and a Long Decode Continuation feature that can stretch reasoning across follow-up calls. For now, though, the model is only rolling out to selected users, and the discount has no confirmed end date.
My take — AI-written commentary, not fact-checked reporting
This is the kind of model release that quietly matters more than the splashy leaderboard brag. A chatbot that guesses less and charges less is the sort of boring win AI vendors love to bury under benchmark confetti. The industry keeps rewarding confidence; users usually just want fewer made-up answers and a bill that doesn’t sting.
Read more about this at: Artificial Analysis