Gemini 4 Argon Hallucinates 15%. Its Accuracy Drops 13 Points.

Ask Gemini 4 Argon something it doesn't know, and it tells you. Ask GPT-6 Astra the same question, and more than half the time you get a confident answer that isn't true.

That's the real story hiding under Tuesday's benchmark tie. Artificial Analysis ran its Intelligence Index on Google's new flagship and came back with a dead heat: Gemini 4 Argon scores 53, matching GPT-6 Astra. Every outlet is running that as a comeback headline. The number that matters is underneath it: on AA-Omniscience, Argon has a 15% hallucination rate, the lowest of any model in its class, compared with 51% for GPT-6 Astra and 54% for GPT-6.1 Sol.

Google didn't build a smarter model. It built a model that knows what it doesn't know.

Here's what nobody wants to sit with: Gemini 4 Argon scores 50% on accuracy, down 5 points from Gemini 3.1 Pro Preview and 13 points below Astra's 63%. That gap isn't rounding error. It's one in three hard questions where Astra lands the right answer and Argon simply declines to guess. Refusing to lie is not free. You're trading wrong-and-confident for right-less-often, and a leaderboard tie can't tell you which failure costs more when a contract closes at 2am.

The sale price is not the business model

At launch, Gemini 4 Argon costs $1.99 per Intelligence Index task, 60% of Astra's $3.26. That discount is a promotion with no expiration date. On standard pricing, Argon's cost per task rises to $3.98. Once the promo lapses, Argon costs about 20% more than Astra, not 40% less. The "60 percent of the cost" framing is a snapshot of a sale, not a pricing strategy Google committed to.

You can't use it yet, and Google explained why

Argon isn't rolling out to developers or paying customers first. Google is giving access to selected cybersecurity defenders through its Fairwind Program, with broader rollout only after defenders have patched system vulnerabilities the model identifies.

Translation: the people who find security holes get first crack at this model, before anyone with a credit card.

That's defensible. It also means the honesty-over-accuracy bet hasn't faced a real commercial customer base yet. "Selected users only" is a safety gate, not a verdict on demand.

For workflows where a wrong answer costs real money—legal review, financial modeling, security scans—a model that says "I don't know" is worth more than one that fills the gap with something plausible. For everything else, Astra's extra 13 accuracy points might just win on output alone.

What to watch

Whether Argon reaches Google AI Ultra subscribers and paid API customers on what timeline. Whether the $3.98 post-promo price holds against Astra once Google stops subsidizing it. Whether OpenAI trains Astra to refuse more often, dropping its own accuracy in the process. Whether Fairwind's cyber defenders surface something in the coming weeks that explains why Google wanted trusted eyes on this model first.