Google unveils Gemini 4 Argon

Investing.com -- Google unveiled Gemini 4 Argon on Wednesday, its next-generation flagship artificial intelligence model, following weeks of signals that the long-awaited system was nearing release. Alphabet shares rose about 2% in after-hours trading following the announcement.

Google CEO Sundar Pichai said in a post on X that there had been “lots of discussion” about the company’s next model and that he wanted to give an early look “as soon as possible.” He introduced Argon as a model showing frontier performance in complex workflows, cyber defense and software engineering.

The announcement follows comments last week from Google DeepMind chief Koray Kavukcuoglu, who said Gemini 4 was in post-training and that Google intended to release an early version “as soon as possible.” He said the company hoped to launch it “much earlier” than the end of the year.

Google has now delivered that early look, but it has not provided a timetable for general public availability. Reuters reported Wednesday that Argon is initially being provided to select cybersecurity partners.

The model is being positioned as a major step up from Google’s previous flagship systems. In benchmarks published by Google, Argon led GPT-6 Astra and Anthropic’s Claude models on most of the tests shown.

Argon scored 68.9% on the Vals Index, compared with 63.1% for GPT-6 Astra and 67.0% for Claude Opus 5.5. It also led on AutomationBench, the Vals Finance Agent benchmark and Harvey’s Legal Agent Benchmark.

Coding was another strong area. Argon scored 77.9% on DeepSWE v1.1 and 91.9% on Vibe Code Bench, ahead of the competing models. GPT-6 Astra, however, led on FrontierSWE v2, while Claude Opus 5.5 posted the highest score on Terminal-bench 4.0.

Argon also showed a particularly large advantage in Google’s long-context testing. It scored 99.7% on GraphWalks at up to 128,000 tokens and 84.2% on the 256,000-to-1 million-token test, compared with 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5.

The model led Google’s tests for science and mathematics on LABBench 2 and RiemannBench, as well as multimodal tests covering charts and video. It also scored 39.5% on Agent’s Last Exam, ahead of GPT-6 Astra’s 34.2%.

The results do not amount to a clean sweep. GPT-6 Astra led several individual tests, including Terminal-Bench Science and OSWorld, while Claude Opus 5.5 topped Terminal-bench 4.0 and PostTrainBench.

Still, the breadth of the results is notable. Google is presenting Argon as a general-purpose system designed for long-running, complex workflows across coding, finance, legal work, science and cybersecurity.

The announcement comes after months of delays and as Google attempts to keep pace with rapidly advancing models from OpenAI and Anthropic. For now, however, Argon’s announcement is more of a preview than a full public launch: Google has demonstrated the model’s capabilities and begun giving it to selected partners, but has not said when consumers and developers will get broad access.

The benchmark results come amid questions about how Gemini 4 performs outside controlled testing. Bloomberg reported that while Gemini 4 has performed well on industry benchmarks, it has been less impressive when employees have used the model for real-world work, citing people with direct access to the effort. Google has pushed back on that characterization, pointing to its internal use of Argon for tasks including coding, quantum-computing research and data-center optimization. The company says the model is capable of handling complex, multistep workflows and has been tested with cybersecurity partners ahead of a broader release.

Original Article

You May Also Be Interested In

• Google: Introducing Gemini 4 Argon
• Logistic Properties of Americas stock surges on Peru asset sale
• Jefferies names top pick in India medical supplies sector