A new model that lets you spend a full 1M tokens across input, thinking, and output just arrived, complete with a developer guide. Title: Gemini 4 Argon: Developer Guide URL: 📝 Overview This is the developer guide for Google DeepMind's latest frontier model, Gemini 4 Argon. It features a unified 1M-token context window and three Thinking Levels to tune reasoning depth. 🧠 Technical highlights Input, thinking, and output now share a single 1M-token window, and Long Decode lets a single response generate up to 1M tokens. The model ID is gemini-4-argon, and prompts over 4,096 tokens get automatic implicit caching at a 95% discount. 📊 Benchmarks Argon ranks #1# on the Vals Index across finance, coding, law, and tax; scores 87% on the robotics benchmark ERQA; and holds attackers to a 0.7% success rate on the Gray Swan IPI prompt-injection benchmark. 🏭 Real-world results Chrome used it to shrink a legacy code-separation effort spanning 6 million lines from months to weeks, and Gemma 4 E2B's inference throughput improved 2.85x. 💰 Availability Available now via Google AI Studio and the Gemini API, at launch pricing of $2 input / $10 output per million tokens (through November 30, 2026). #Gemini4Argon# #DeveloperGuide#

gemini 4 noticeably different from other models when it comes to refusals and surprisingly few people are talking about it it's noticeably less sensitive than most models it still draws the line at clearly harmful requests but outside of that it's willing to talk about very much anything

Gemini 4 Carbon FPS Zombie Game This was 1 shot, took around 30 minutes to generate everything from scratch This genuinely makes Argon look horrible and I find that very intriguing

Gemini 4 Carbon is absolutely crazy This is the 3D seaplane prompt on it, it created an infinite ocean with randomised island spawns, plus driving controls The UI problems still need fixing but this was quicker than the previous checkpoint

Gemini 4 Argon combined three open-source datasets to build an interactive 3D visualization of human anatomy. Toggle different layers in motion, isolate muscles, or zoom inside a beating heart.

Gemini 4 Carbon is incredible at making games Checkpoint 1 is testing in arena currently and this is one of the cleanest outputs I have got from the Super Mario prompt Gen time was quite high like 30 minutes for this but it's no longer lazy like Argon, it just gets what you want and does it perfectly first try

Gemini 4 Carbon Voxel Pagoda Carbon has been in arena since yesterday, so I will show off some outputs I got over today It's a noticeable improvement over Argon but is incredibly slow, like 2-4x slower depending on task

Gemini 4 Carbon EAP has officially started. Early access has been given to partners and it's being tested in LMArena. The release is very close. It's noticeably better than Gemini 4 Argon at 3D, though frontend/UI is still bad. It's also slower but less lazy.

gemini 4 scores horribly on FrontierMath Tier 4 with openai models already saturating this benchmark, gemini 4's 78% is way below what i expected where is google's deep think? they were making huge progress in research-level math months ago yet we still haven't seen what it can do on this benchmark