Gemini 4 makes the cost of intelligence worth another look.

There is a catch in the price comparison that matters just as much as the headline. Google's introductory API pricing is $2 per million input tokens and $10 per million output tokens. Google says those rates will later rise to $4 and $20, and Artificial Analysis estimates that the model's cost on its Intelligence Index would rise to $3.98 per task at the standard rates. It also found that Argon used more output tokens per task than Astra in this evaluation. In other words, the initial advantage comes from the price of the tokens, not from Argon doing the work with fewer of them. A company should compare the cost of a finished, acceptable task, including retries and review, instead of assuming the cheaper rate on a pricing page will always produce the cheaper workflow.

The other result that stood out to me was Artificial Analysis's 15% hallucination rate for Argon on AA-Omniscience. Among models scoring at least 45 on its Intelligence Index, it was the lowest rate the evaluator reported. The test asks factual questions and looks at what a model does when it cannot give a correct answer. Its hallucination measure is the share of those non-correct responses that are wrong guesses, rather than partial answers or admissions of uncertainty. On that measure, Argon was much more likely to say it did not know than to make up an answer. That is a useful behavior when someone is relying on a model to research an issue or find information in company documents.

That 15% does not mean only 15% of all Gemini responses are wrong, and it does not mean the model will always admit uncertainty. Artificial Analysis reported 50% overall accuracy for Argon on the same knowledge test, below the 63% it measured for GPT-6 Astra. There is a tradeoff here: a model that is more willing to stop and say it does not know can save a team from acting on a confident invention, but it may also leave more questions for a person to resolve. For a business, both sides matter. An honest gap is easier to work with than a fabricated answer, provided someone owns the next step.

Argon also leads Zapier's AutomationBench, which tests agents completing work across business tools, at about 51.3%. That result shows real progress and a clear limit: even a leading model does not complete every multi-step process correctly. The question for a smaller company is where a better, less expensive model could help and where the workflow still needs reliable source information, a review point, or a person making the final decision.

This is why Ruby begins with the work rather than the model announcement. We sit with employees to understand the questions they answer repeatedly, the information they need, and the parts of a process that take too much time. From there, we can test whether a tool makes that work more accurate or faster and measure what it costs to get an answer the team can actually use. As platforms change, our ongoing support helps employees understand which updates matter to their workflows instead of asking them to track every new release themselves.