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

The poster says the previous day's Gemini 4 Pro leak was pointing at Argon, and after the official release argues Google is back in the AI race.

Argon supports 1M output tokens versus 128K for Astra and Opus 5.5, scores 77.9% on DeepSWE long-horizon software engineering against 74.1% and 74.2%, and leads by 8-10 points on business workflows, financial research, legal tasks and long-video understanding.

An analysis says Argon costs $1.99 per task versus Astra's $3.26, but that reflects a half-price introductory rate; at Google's list pricing it works out to about $3.98, roughly 20% more than Astra.

A Japanese user says Gemini 4 Argon has launched with top-tier scores on major benchmarks at about 60% of Astra's cost, though it is currently limited release only.

First-day feedback welcomes its long-context and security positioning but questions Google-reported scores and complains that regular users cannot access it yet.

A post claims Argon beats GPT-6 Astra and Opus 5.5, listing a 2M token context window, up to 1M tokens of output and a 15% hallucination rate.

An Indonesian post lists internal Google uses for Gemini 4 Argon, including clearing 300TiB+ of data center memory, rewriting 800,000 lines of C/C++ kernel code in Rust and a 2.7x faster video decoder.

A post says Google has finally released its frontier model Argon after a run of Flash models, with first-place benchmark results that beat Opus and Astra in some tests, though most people can't use it yet.

The poster confirms Gemini 4 Argon has been announced with impressive benchmarks but says they care more about how it feels to talk to and will write a review once it's released.

A post argues that raising Argon's output ceiling from 64K to 1M tokens gives agents room to run full engineering loops, citing Google internal work on Fuchsia kernel migration and libgav1 video decoding.

A video post claims Gemini 4 Argon outperforms GPT-6 Astra while costing about 80% less, without specifying which tests were run.

The author says they gave AI agents GPUs to train better AI, but Gemini 3.8 Flash stopped training after about 12 hours, and publicly asked Google for Gemini 4 to take over.