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

A Japanese user says they are tired of things like Mythos and wants Google to release Gemini 4 Pro sooner.

One user argues Gemini 4 is open to only a few people, letting Google claim a lead over OpenAI and Anthropic before wider testing can confirm it.

A Spanish creator says Google's next generation, Gemini 4, uses new names like Argon and beats Astra and Fable on several benchmarks.

Google announced a new model, Gemini 4 Argon, claiming it outperforms competitors like Claude and ChatGPT on several important benchmarks.

A Spanish-language post expresses surprise that Google had Gemini 4 Argon in reserve, saying it competes evenly with Astra and Fable and surpasses them on several benchmarks Google published.

Comparison figures put Gemini 4 Argon (high) at a 15% hallucination rate on AA-Omniscience, far below Astra at 45-53% and Opus 5.5 at 59-66%.

A reply argues that Gemini 4 Argon at high reasoning exceeds Astra and Opus at high reasoning in intelligence, so its cost-to-performance ratio remains good.

The author rendered two short films with Minimax to compare the creative output of Gemini 4 Argon and Gemini 3.8 Flash.

The author compares $2/$10 Argon, $4/$20 Opus 5.5 and $10/$50 Astra, arguing Argon's low token price is misleading because it burns more tokens to finish the same task.

LuminaBench links a paywalled Bloomberg report on Google employee skepticism toward the new Gemini model and says its own testing shows Gemini 4 Pro is a frontier model.

A Portuguese-language user says Gemini 4 Argon is not as efficient as GPT-6.1-Sol but posts strong benchmarks, priced at $2/$10, and plans to test it in Antigravity.

A roundup says Gemini 4 Argon has the lowest hallucination rate, tops the Vals index and text arena, scores 53 on AA to match Astra, and costs less per task but is not token efficient.