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

Migma AI announced on October 14 that Gemini 4 Argon is live on its platform, turning a single brief into branded emails, paired with Mailcheap sending and click analytics.

StatsWire says Gemini 4 Argon beats GPT-6 Astra and Claude on heavy data analysis, citing a 15% hallucination rate, 68.9% on the Vals Index, 65.4% on Finance Agent and a 1M context window.

After Gemini 4 Argon's release, someone gave it $10k to trade live markets against Claude, ChatGPT, Grok, DeepSeek and Muse to compare returns.

WesRoth says the highlight is not the seven straight throws but that Argon can build its own practice simulation, which could make teaching robots new tasks much cheaper if it holds up across different objects, distances and scenes.

Someone notes that Argon's 78% on Frontiermath Tier 4 is not impressive by current frontier standards, and remarks on how fast models have advanced since the start of the year.

On the HSS sixth-sense test, Argon scores 55.4% overall against Astra's 53.6%, leads 59.1% to under 41% on clips over ten minutes, but loses to Astra on short videos.

A former Gemini team member congratulates the Gemini team on Gemini 4 Argon taking first place on the Humanity's Sixth Sense visual commonsense reasoning benchmark.

A post citing Sergei Bren says Google is using RSI loops on Gemini 4 Argon, compares the effort at Grok, OpenAI Astra 6.1 and Meta Muse Wonder, and asks why Anthropic is still ahead.

A 30-minute Google for Developers talk covers internal testing by 100,000 Googlers, 3-4 week cycles, using AI to check AI, and what 1M output tokens are good for.

A tally of 400 stories per model lists each model's recurring phrases, with Gemini Argon often using hearts hammering against ribs, reaching into coat pockets and the dark water.

Comparing Gemini 4 Argon and Opus 5.5 on the same Chun-Li Blender test, the poster says Opus still holds the edge on detail.

Google had Argon write 5,000+ lines of zero-dependency Python to build its own ray-tracing engine, video tool and audio mixer, then generate narration and music.