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

A Vietnamese tech roundup says Google widened Gemini 4 Argon access for deep tasks, pits it against Claude Opus 5.5, and notes a firm appointing an AI strategy head.

A post with a screenshot claims Gemini 4 Argon is available in AI Studio for both free-trial and paid Pro users.

First-round Antigravity testing finds Gemini 4 Argon's coding quota on Gemini AI Pro is roughly 1/17 of Opus 5.5 on Claude Pro and 1/7 of GPT-6 Astra on ChatGPT Plus.

A user finds Argon poor in the Gemini app yet excellent in AI Studio, suspecting the harness or a newer checkpoint.

After days of heavy use, the author rates Gemini 4 Argon below Claude and GPT on intelligence, flags heavy sycophancy, missing work-mode and project features, and tight usage limits.

A user complains that Gemini 4 Argon's Pro plan limits are so tight that four prompts exhausted the entire 5-hour usage cap.

A user spotted Gemini 4 Argon newly added to Google's grouped model listing, suggesting a possible rollout or limited access.

The author argues Argon is slow and expensive with only small gains over 3.8 Flash, and that Gemini 4 Flash should build on Argon's foundation while being cheaper, faster and better at coding.

A Japanese post rounds up Gemini 4 Argon's selling points: 1M-token context and multimodal support, with Google-reported scores of 68.9% on professional tasks and 77.9% on coding, both above GPT-6 Astra and Claude Opus 5.5, though Claude is stronger at on-device operation.

Ammaar used Gemini 4 Argon to build a Roadster racing game playable in the browser; an auto-run on a phone in portrait scored 5,475 points in 30 seconds, and the author says the 3D and sound reach KartRider quality.

Venice has added Gemini 4 Argon with anonymous access, priced at $2.50 per million input tokens and $12.50 per million output tokens.

A prompt lets Grok Bot connect to Gemini through the Antigravity CLI with the Argon model as default, after signing in to a US Google account in the cloud browser.