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

Artificial Analysis says Gemini 4 Argon (high) matches GPT-6 Astra (max) on its Intelligence Index, with a 15% hallucination rate versus 51% for Astra.

A Spanish post claims Gemini 4 Argon writes responses 8x longer than rivals, performs about 5x better than Opus 5.5 on legal tasks and costs half at launch, but is limited to a small group of cybersecurity experts for now.

A video post says Gemini 4 Argon is priced below GPT-6 Astra and Opus 5.5, pitching cheaper high-intensity reasoning.

At its highest reasoning setting, Argon scores 53 on the Intelligence Index, one point above GPT-6.1 Sol and 23 above Gemini 3.1 Pro Preview; its hallucination rate is just 15%, but its accuracy of 50% trails Astra's 63%.

Having found Gemini 3.8 Flash strong at Japanese expression, the author hopes Gemini 4 Argon will soon reach Google AI Studio so users outside the pricey Ultra plan can try it.

Google released Gemini 4 Argon, raising the output token limit from 64K to 1M, with DeepSWE reportedly matching Opus 5.5 and Astra, and early access going first to cyber defense organizations.

The author made a short film about an underwater library collapsing and its last librarian, a cinematic take on the Gemini 4 Argon launch set against Gemini 3.8 Flash, created with Medeo and Seedance.

Bloomberg sources say Gemini 4 is very large and costly to run, and two people familiar with it say it may be shaped by an industry focus on benchmarks; Google says it is limited to trusted testers and will first go to paid API customers and AI Ultra subscribers.

A poster says the Gemini 4 Pro rumors were true, that it leads GPT-6 Astra and Claude Opus 5.5 on benchmarks, and jokes about OpenAI's $500 plan.

Citing figures like 77.9% on DeepSWE, first place on the Vals Index and up to 1M output tokens, the author says that if they hold, the contest shifts from OpenAI vs Anthropic to a three-way race that includes Google.

On the new Terminal-Bench-Science benchmark, Gemini 4 Argon jumps from 12.4% for Gemini 3.8 Flash to 57.6%, placing third on the leaderboard.

A Korean post claims Gemini 4 Argon benchmarks on par with Opus 5.5 and Astra 6, with a 1M context window, $2/$10 per million input/output tokens, a 95% cache discount, and Ultra subscribers first.