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

A pricing comparison puts Kimi K3.1 Pro fourth globally and first in China, cheaper only than Fable 5.1, Opus 5.5 and GPT Astra, but pricier than Gemini 4 Argon, GPT 6.1 Sol and GLM 5.3.

The post lists Argon aiding plasma physics, five lab-verified enzymes, a missed newborn genetic abnormality and near-doubled quantum error-correction speed at over 100x lower cost.

A user who bought a cheap Pro membership finds Gemini 4 Argon only runs in Spark Beta with no model choice, hits version-number hallucinations and burns through usage limits fast.

A long post claims free Pro accounts get routed to a cheaper gemini-4-argon-trail branch while Ultra gets the original, and Antigravity itself runs at about 80% of AI Studio/Vertex quality.

The post cites Pichai saying Google is capacity-constrained amid the Argon model launch, using it to push back on AI bubble talk.

A Google blog post showcases projects built with Gemini 4 Argon by several developers and LuminaBench, linked in the post.

Long post analyzes Kimi K3.1 Pro's post-training gains, 65.7% on Terminal-Bench 4.0, roughly half of Opus 5.5's price, and an Artificial Analysis score of 55 that leads GPT-6 Astra and Gemini 4 Argon.

A reply says pairing Gemini 4 Carbon with Gemini 4 Flash would be a top-tier combo, mentioning a Carbon variant rarely discussed.

A benchmark shows Kimi K3.1 Pro scoring 9% on real-robot tasks, trailing Opus 5.5, Gemini 4 Argon and GPT-6 Astra.

A developer says Gemini Argon may be the first LLM whose actual runtime matches its stated estimate, though every task still feels slow.

A user says free Google AI Pro access no longer works, so they paid $10 on Gamsgo for a 3-month Pro subscription, believing the enterprise-invite route reliably unlocks Gemini 4 Argon, and shares an invite code.

A Korean post claims Kimi K3.1 Pro surpasses GPT-6 Astra, Fable 5.1, Gemini 4 Argon and Muse Wonder, with a big jump in agent performance.