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

A developer counters claims that Argon can't write code, saying it split a 4,290-line monolithic HTML file into modular ES6, switched to glb streaming and optimized GPU buffers with no visual change.

A Vietnamese post says Google's October 13 update brings Gemini 4 Argon to AI Pro and Ultra users worldwide, targeting complex tasks like report analysis, slide building and 3D modeling.

A user explains why they keep asking for GPT-4o's return, citing its rare depth of intelligence and empathy, and says this holds even with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini Argon available.

Zhipu's GLM-Legal ranks first among open-source models on Vals AI, passes 19.2% of Harvey LAB tasks, and costs $0.15 per task, about 70x less than the similarly scoring Gemini 4 Argon.

The author argues the score measures the model plus the benchmark runner, including tools, time limits and retries, so 21% does not necessarily mean the model is weak at coding.

A roundup of rumored upcoming models: OpenAI Rosalind, Anthropic Fable 5.5 and Iggy, Grok 4.8, Meta Watermelon, and Google's internal Carbon.

A post with an image asserts that Argon is not a Pro model, suggesting Gemini 4's tier lineup differs from rumors.

The poster shares a video comparing Gemini 4 Argon with Muse Wonder, with text only calling the results insane and no description of what was tested.

The author says prompt-guided Argon has completed about 75% of a WebGL test app, with the author only specifying features and reporting bugs, and promises to share the full process once it is finished.

A Japanese outlet rounds up five Argon updates: an offline-capable task mode, 335 hours of work done in 7 days, Google-internal data from 100,000 employees, and Antigravity integration in the Mac app.

A short positive note says the combination of Gemini Argon and Deep Research has been incredible, without giving specific test details.

Given the same "3D human anatomy atlas" prompt, Gemini 3.8 Flash returned floating dots while GPT-6 Astra built an interactive atlas of muscles, bones and organs.