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

Students report tiny refunds on wiped prepaid balances, flagged re-subscriptions, and Gemini 4 Argon absent from new student plans in Antigravity.

The author argues attention on Gemini 4 Argon benchmarks misses Google's bigger plan, possibly Carbon for coding and Flash for scale.

A reply says the hallucination problem was supposedly fixed in Gemini 4 Argon, implying it still fabricates answers.

A joke about release timing: Fable 5.5 and Astra 6.1 each tell the other to go first, Google says it will hold Gemini 4 Carbon a bit longer, and DeepSeek is left out.

The post contrasts Gemini 4 Argon's 1M output tokens with Opus 5.5's 128K and asks whether this is a real edge for long code migrations or just marketing.

After a purge of users abusing free access, Gemini 4 Argon Pro's quota has reportedly risen sharply, slightly topping ChatGPT Plus, with hopes Antigravity's Claude quota follows.

Antigravity ACP References 1.4.3 reportedly lists gemini-4-carbon, carbon-g-medium-tf and Argon max-mode/teamfood, suggesting internal teamfooding has begun and a launch as early as next week.

After a day testing Gemini 4 Argon, the poster suspects Google's heavy Argon promotion hides a weaker Carbon, and doubts a model surpassing Opus 5.5 is near.

Google CEO Pichai says RSI is a critical contributor to developing Argon, but today it's still RSI loops alongside researchers with agents, far from the end state.

A short travel video originally made with Seedance 2.5 was recreated by its author using Gemini Argon.

A Japanese user compares model answers on a Keio prep mock exam: Gemini 4 Argon agrees with Claude Opus 5.5, while GPT-6 errs on world history.

A user reports that free and Pro tiers can occasionally be routed to a model named gemini-4-argon-trial, likely a Gemini 4 Argon trial rollout.