Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

Google DeepMind shows Argon deriving physics from a video, building its own simulation, and teaching Apptronik's Apollo humanoid to sink 7 throws in a row.

A Turkish user says Gemini 4 Argon on the Ultra plan ran out of its 5-hour limit after just 4 prompts, and wonders how much worse the Pro plan would be.

A user had Claude Opus 5.5 make an intro video for Gemini 4 Argon, saying it looks more presentable than Argon's own and joking Google should outsource its launch videos to Claude.

OpenCode announced that Gemini 4 Argon is now available, noting it was the team's second most-used model last week.

A one-line opinion arguing that Google has returned to the model capability frontier with Gemini 4 Argon, with no benchmark details.

A screenshot shows Gemini 4 Argon appearing in the Gemini app, available only to Ultra subscribers.

A video says Gemini 4 Argon is live, scoring on par with GPT-6 Astra in third-party benchmarks, with improved agent capabilities, the lowest hallucination rate among 45+ models, and 50% off until November 30.

A user says Gemini 4 Argon is now available on fx through AI Gateway at 50% off until December 1, with a video demo.

Factory announced day-0 support for Argon, reporting it fixed input validation without breaking existing functionality and delivered a multi-module idempotent app at high effort.

A Turkish post says Gemini 4 Argon opened today, raising single-response output from 64K to 1 million tokens and topping the Vals Index at 68.9%.

A post says Google has made Gemini 4 Argon available worldwide and suggests two agent ideas: an auditor that fixes existing agents and an enterprise agent that runs ops workflows end to end.

A Japanese post with a screenshot announces the general release of Gemini 4 Argon.