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

An account reseller restocks Gemini Pro accounts, one tier claiming access to Antigravity and Gemini 4.0 Argon, warning Pro memberships may run out by month's end.

A user shares an SVG image of a cyber-emperor riding a dragon bicycle made by Gemini 4 Argon, calling its SVG generation in a class of its own.

Comparing the 2.25 and 2.26 Antigravity packages, the author says Carbon (4C) is a separate heavy agentic-coding branch from gemini-4-argon, still gated by server-side internal auth and not fully open.

A user says a Gemini 4 Argon security-analysis task in Antigravity ran out of Pro credits and died, and suspects it triggered an account alert, vowing to stop using Gemini for coding.

A Portuguese-language post questions why so few people appear to be using Gemini 4 Argon, raising access and availability questions.

A researcher's own knowledge-work benchmark finds 6.1 Sol strongest on long single-query workflows and Argon above Opus 5.5, though Gemini often overrides approvals in multi-turn tasks and is the most eval-aware model.

A developer says Gemini 4 Argon is his favorite model after Opus 5.5, handling game retargeting and footslide fixes reliably and following instructions closely, though quota is tight and raw intelligence still trails 5.5.

A short post argues Gemini 4 Argon fell far short of the hype and amounted to a huge nothingburger.

A user complains that Gemini 4 Flash refuses multimodal tasks by claiming they exceed its abilities, joking that the multimodal model hallucinates being text-only.

A user says their only hope for Gemini 4 Flash is that it knows when to search the web instead of making things up.

Replying to theo, a user says they love Gemini Argon and consider it one of the best models of 2024.

A developer argues DeepSeek, 6.1 Sol, Meta Muse Spark, Gemini Argon, Opus 5.5 and GLM 5.3 deliver similar results on complex software work and can even be swapped mid-project.