Google rolls Gemini 4 Argon into its app for Pro and Ultra subscribers

The tasks option gives subscribers another way to use Gemini in the app. Google says they can switch from chat to tasks and leave while Gemini continues working. The company says Gemini connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, and presents the integrations as supporting activities such as grocery ordering, booking tables and design work.

Those are Google’s descriptions of the feature, not independent results showing how often it completes tasks correctly. The announcement does not establish how reliably tasks finish or how often they need user input.

Google’s announcement also makes Argon available to developers and enterprises through AI Studio and the Gemini platform. Google AI Ultra subscribers can use it in Google Antigravity, the company said.

What evidence supports Google’s problem-solving pitch?

Google presents Argon as a general-purpose frontier model for complex work, including coding, finance, legal work, cybersecurity and science. Its examples of engineering, scientific and security results are reported by Google and its partners. The company’s overview does not independently verify those results or establish that Argon can reliably solve difficult problems across those fields.

Google DeepMind has also published robotics prompts and videos showing Argon in robot-control and simulation scenarios. Those are company demonstrations, not outside evaluations of how well the model controls robots. DeepMind reports an 87% score on its Embodied Reasoning benchmark, but the supplied material does not establish independent replication of that result.

External benchmarks offer a separate, more limited measure. Artificial Analysis gives Argon a score of 53 on its Intelligence Index; its comparison with GPT-6 Astra gives that model 51, while showing different results across the component tests rather than a uniform Argon lead. Vals AI lists Argon first among 41 models on its Vals Index, with a score of 68.90%.

Those scores are evidence about Argon’s performance on the benchmarks each provider measures. They do not show that it is best at every task, confirm Google’s reported business outcomes or guarantee that it will solve an individual user’s difficult problem. A ranking is not a substitute for testing a model in the work where someone plans to use it.

What can users verify for themselves?

Users can check particular outputs when a task has a verifiable result. For example, code can be run against tests, calculations can be checked, and factual answers can be compared with reliable records. Such checks help assess an individual response; they do not establish how reliably Argon performs across other prompts, fields or workflows.

The rollout therefore changes who can try Argon, not what has been independently established about its broadest claims. The pace at which access reaches eligible app accounts remains unclear, and the supplied material includes no independent replication of Google’s reported case studies or Embodied Reasoning score.