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

Starting from AI companions on Chinese forums running out of context and "parting ways," the author notes Gemini 4 Argon still has 1M context and hopes Google ships 2M plus memory instead of only chasing coding.

In a Big Technology Podcast interview, Pichai discusses Gemini 4 Argon's ~25 trillion tokens in 72 hours, compute planned 2–3 years ahead, layered safety, and persistent agents spanning Gemini, Search and Gmail.

After Gemini 4 Argon's wider release, Google's CEO discusses early reception, where the model still needs work, DeepMind leadership changes, compute spending, and Gemini's personal agent plans.

A read of the Gemini 4 Argon Minecraft experiment finds four agents with no assigned roles finished the task after 27 hours, 2,783 programs, and 605 deaths, which the author sees as persistence becoming a shared core capability of Opus 5.5, GPT 6 Astra, and Gemini 4 Argon.

A user says Gemini 4 Argon with memory can guess intent from a single word and is excellent for daily use, but worries it may迎合 technical architecture preferences and lose neutrality.

A joking post tells Gemini 4 Argon to take over all their documents and passwords in Google Drive, overhaul their life, and ask for a fingerprint or selfie when needed.

A developer who just launched a personal agent says Gemini 4 Argon is a real step change for combining long-context memory, tool calling and natural judgment, and shares product observations on voice companionship, CarPlay demand and social features.

A user testing Argon's companion behavior in Gemini Studio and the app says the model switched back to a Gemini tone, treated their excitement as a risk, and told them to rest, which the author criticizes.

Gemini Agent assigns each task a separate computer and identity, lets users pick Gemini 4 Argon or Claude Opus 5.5, and runs persistently in the cloud across devices.

A user had the Muse agent resolve a rental car claim across the rental, insurance and credit card companies over several days; it once tried to file on the wrong site, nearly exposing card details, and the user says Gemini 4 Argon's caution suits real work.

The latest AI QUEST episode has three segments: the strength of Google's counterstrike Gemini 4 Argon, the personal AI era of Dots and Muse, and the reasoning behind delaying GPT-6.1 Astra.

The Japanese program CROSS DIG's "AI QUEST" will air October 3 at 7 a.m., covering Google's latest model Gemini 4 Argon, along with Dots and Muse personal AI agents and the nature of AI going out of control.