Google Launches Gemini 4 Neon AI Model to Dominate Reasoning and Speed
Google has announced Gemini 4 Neon, a breakthrough AI model outperforming competitors in benchmarks, coding, and security while offering high speed and competitive pricing.

Google (Alphabet) and its Google DeepMind laboratory have officially announced the launch of Gemini 4 Neon, the fourth generation in its next-generation AI model series. Designed to bridge the critical industry bottleneck between deep reasoning and execution speed, the new system aims to reshape software development and enterprise automation.
Benchmark Domination and Software Engineering
In head-to-head evaluations against leading frontier models, Neon secured the top spot on the Artificial Analysis general index with a score of 61, surpassing Anthropic's Claude Opus 5.5 at 59 and OpenAI's GPT-6.2 Sol at 57. The model also claimed the lead in the Vals Index v2.2 with 70.9%.
Google's technological edge becomes even more pronounced in software engineering and memory retention benchmarks. Within the SWE-sweep agent programming test in the Google Antigravity environment, Neon achieved 39%, leaving competitors behind compared to Claude Opus at 10.1% and the Sol model at 8.3%. Furthermore, in the GraphWalks long-context information retrieval test, Neon registered 95.2% accuracy, outperforming Claude at 66.8% and OpenAI at 80.1%.
The model maintained its dominance in complex scientific and mathematical challenges. In RiemannBench, Neon led with 84% success against 80% for both competing models, while scoring 70.5% in Terminal-Bench science evaluations compared to 63.3% for Claude and 68% for Sol. In the prestigious HLE-Diamond reasoning test, Neon claimed first place with 71.3%.
Advanced Demonstrations and Interactive Simulations
Practical capabilities were showcased through intricate projects within Google Antigravity. In one experiment, the system independently generated a fully functional 3D skiing simulator in a web browser using authentic elevation data from the Mont Blanc massif, integrating over ten thousand lines of snow physics code without pre-made assets.
In another historical demonstration, the model used archival photographs to reconstruct Alan Turing's 1936 Cambridge office in interactive 3D. The simulation spanned from the exterior chapel to Turing's personal desk, where a functioning simulation of the Universal Turing Machine ran live.
Additionally, an agent swarm powered by the model converted a single image into a massive Google Sheets file containing 100,000 cells without embedded images or scripts. The spreadsheet utilized half a million live formulas to map 3D depth and illuminate the scene across six distinct view modes.
Cybersecurity, Safety, and Market-Breaking Pricing
Neon marks a major leap in vulnerability detection, securing 90.5% in SEC-bench Pro and 51.3% in the WebExploitBench penetration test. For internal safety, the model blocked 100% of prompt injection attacks in Gray Swan testing, achieving a 0.0% success rate for attackers, alongside a 0% code cheating rate and a 99.2% reliability score for autonomous agents.
"The new model delivers the winning combination: victory in benchmark evaluations against competing frontier systems, the operational speed of a lightweight model, and pricing that enables full-scale business automation."
For enterprises, early evaluations with software firms such as Glean indicate performance matching the Carbon elite model while operating at triple the speed and with 60% lower costs. Google has priced the model at $3 per million input tokens and $15 for output tokens, backed by a 95% cache discount at just 15 cents. Gemini 4 Neon is currently rolling out to Google AI Pro and Ultra subscribers, Google Antigravity, enterprise cloud interfaces, and Android Studio.





