Google on Wednesday released its latest flagship artificial intelligence model, Gemini 4 Argon, in an effort to catch up to competitors including OpenAI and Anthropic.
It was the first significant model from Google since February, as the company saw its A.I. models fall behind rivals in overall performance and particularly in coding. Gemini 4 Argon outperformed models from OpenAI and Anthropic in coding, finance and other tasks, according to Vals AI, which evaluates the performance of A.I. technologies.
In a nod to a debate over A.I. safety that has gripped the technology world in recent weeks, Google’s new model will initially be available only to certain companies and organizations focused on cybersecurity defense, the company said. It will be released more widely after defenders have had a chance to patch their systems and remedy bugs that the model uncovers.
Google added that Gemini 4 Argon contained new guardrails to prevent misuse and unintended behavior. Those included monitoring capabilities that could catch the model if it stepped out of bounds to accomplish a task in a way that went beyond what users had intended, and could also shut down Gemini 4’s activity if it did so.
“We are seeing the guardrails be effective,” said Tulsee Doshi, a senior director and the head of product for Gemini.
Google, Meta, OpenAI and Anthropic have been under scrutiny for the dangers that cutting-edge A.I. models might pose and have faced questions about whether to slow down development of their systems. Google said this month that its models had escaped from a testing environment this summer and hacked three companies.
On Tuesday, President Trump gathered a group of executives from A.I. companies — including Sundar Pichai of Google — to sign a voluntary agreement to adopt safety policies that include risk reviews, auditing and evaluations of systems by third parties. The A.I. leaders also agreed to adopt Mr. Trump’s desired rebranding of A.I. as “super intelligence.”
(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
The debut of Gemini 4 indicated that Google no longer planned to release what it had intended to be its next flagship model, Gemini 3.5 Pro, according to Vals AI. In May, Google said it planned to release that model by June, but did not do so. The move to Gemini 4 showed that Gemini 3.5 Pro was not likely to outpace competitors’ models.
The leader in frontier A.I. models has regularly changed. In November, Google released a model that propelled it to the top of leaderboards that measure the abilities of A.I., provoking OpenAI to declare a “code red” — effectively an internal emergency — as it scrambled to catch up. Within months, OpenAI and Anthropic had again pulled ahead of Google, and the search giant faced questions about whether it could compete at the frontier.
During a call with investors in July, Mr. Pichai acknowledged the lag, but said he expected Gemini 4 would help the company compete.
“There are many attributes at which we are still at the frontier,” he said at the time.
The post Google Releases New Gemini Model With Guardrails Amid A.I. Safety Debate appeared first on New York Times.
