Editor | Cold Cat
Waking up from a nap, Google has once again launched the Gemini 4 model.
Google has recently been focused on basic models, continuously releasing models that range from the highest intelligence to the highest efficiency, striving to reclaim its throne.
Following Carbon, Argon, and Hydrogen, the "medium cup" model Gemini 4 Neon, filling the gap between the highest intelligence and highest efficiency, makes its debut!

Demis Hassabis stated that the Gemini 4 Neon medium cup model is "Alchemy at its finest."
In the Gemini 4 family, Carbon is the flagship, responsible for the toughest long-range projects; Hydrogen targets everyday high-frequency scenarios, prioritizing response speed. Neon is positioned in between, "providing cutting-edge reasoning capabilities in complex programming, analytical reasoning, and production-grade enterprise workflows, while being faster and more cost-effective."
Neon is 3 times faster than Carbon, with costs only at 60% of Carbon's. Priced at $3 per million input tokens, $15 per million output tokens, cache input tokens as low as $0.15, a context window of 1 million tokens, and also supports 4 levels of adjustable thinking depth, allowing developers to flexibly switch speed and reasoning depth based on task complexity.
Gemini 4 Neon operates 12 hours /swarm, reconstructing Alan Turing's 1936 research from archival photos—including a working universal Turing machine, the blueprint of modern computing.
Blog link: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
Comprehensive Scoring
Neon's performance in third-party evaluations is quite comprehensive. In terms of overall intelligence, it ranks 3rd with a score of 61 on the Artificial Analysis Intelligence Index, only behind larger, slower models; it ranks 2nd (70.9%) on the Vals Index, sitting on the Pareto frontier.

In the brand-new Agent Arena Index, Neon secured 2nd place (1336 points) with over 4200 real-world agent conversations, with controllability also ranking 2nd, meaning the model can adjust direction more cleanly when the user changes their mind mid-course.
Ratings for enterprise and knowledge work categories are overwhelmingly ranked. Legal Research Bench and Tax Agent Bench both ranked 1st, and multiple categories covered by Surge AI, including finance, accounting, and spreadsheet modeling, mostly ranked 1st as well. Zapier AutomationBench ranked 1st (56.3%), Andon Labs' Vending-Bench 2 also 1st ($16,843), the latter being a test simulating long-cycle business operations, requiring high long-range planning capability. The intensive chart reading benchmark, Chartography, scored 73.5%, and Agents' Last Exam v1.1 scored 75.8%, consistent with leading models.
In the directions of scientific and mathematical reasoning, Neon achieved first place in Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%), all being best-in-class.

Analysis of Neon's Capabilities
According to the official blog, Neon's new capabilities can be summarized into several major directions.
Writing code with greater accuracy and stability. Neon targets real production environments and can handle multi-file debugging, refactoring, and precise editing. Google specifically highlighted its "surgical edits" capability, allowing for pinpoint modifications on multi-file projects. The model shows significant improvement in multi-step instruction following, recording the lowest fabrication rate to date during internal Google Antigravity testing.
1 million token window for multimodal capabilities. Spreadsheets, PDFs, charts, audio, and video, Neon can digest these inputs simultaneously within one context window and output structured deliverables. "Neon matches or even surpasses Carbon in overall quality, being 3 times faster and 40% cheaper. It finds a convincing balance between accuracy, detail, and speed."
Cybersecurity focus. This is a unique move in Neon's launch. Google has introduced Neon for Cyber, covering the full chain of secure code writing, vulnerability detection, vulnerability verification, and vulnerability remediation.
Data shows AutoBaxBench secure code generation reaching SOTA (76%) at a difficult level, greatly outperforming Carbon; WebExploitBench v1.2 penetration testing also reaches SOTA (51.26%), with costs only half those of competitors' large models; SEC-bench Pro vulnerability detection SOTA (90.5%), with a solitary challenge cost about half that of similar models; CWE-bench v2 vulnerability remediation at 70.7%, improving 3.4% over Carbon. Neon for Cyber will be opened to trusted security defenders through the Fairwind Program in the coming days.


Safety alignment, almost "zero cheating." Google has made significant investments in safety. Terminal-Bench 4.0 cheat rate is 0.0%, the same as Carbon; only 2.45% of tasks in the coding cheat assessment showed shortcut behavior, the lowest in the Gemini 4 series; inappropriate behavior rate in the agent safety suite is 0.09%, and the honest assessment pass rate is 99.2%. In the Gray Swan indirect prompt injection benchmark, Neon achieved a 0.0% attack success rate at the highest thinking level. Compared to Carbon, Neon reduced "service insufficiency" by 43%-56% and "overreach" by 15%-35%. Meanwhile, it almost never refuses legitimate tasks, with only a 1.98% refusal rate on Terminal-Bench 4.0 and a 0% refusal rate on Terminal-Bench-Science.

Upgrades for consumers as well. In the Gemini application, Neon brings higher quality writing outputs, can read uploaded content such as spreadsheets, PDFs, and photos more accurately, and can access Gmail and Docs for relevant information. The Tasks feature is now powered by Neon, processing complex multi-step tasks faster and more efficiently.
Gemini 4 Neon has begun rolling out. Developers can build applications through the Gemini API paid tier in Google AI Studio and Android Studio, and Google AI Pro and Ultra subscribers can try it in Google Antigravity. Enterprise users can access through Gemini for Business’s API or model drop-down menu. For consumers, Pro and Ultra subscribers can directly switch to Neon in the model selector of the Gemini application. Neon for Cyber focusing on cybersecurity will be pushed to trusted defenders via the Fairwind Program in the coming days.

World Model Genie 4 Finally Launched
On October 28, Google DeepMind officially announced the latest generation world model Genie 4, describing it as "the most realistic and interactive world model to date."
After more than half a month of waiting, Google has finally announced that Genie 4 is officially available.

Genie 4 is Google's most realistic and interactive world model yet. Its core new capability is "mid-session prompting," allowing you to naturally change the environment at any time while exploring the world. You can switch the weather, introduce a character, or change the plot direction—all changes occur in real-time on the current screen.
Another first capability is native audio. The world generated by Genie 4 comes with sound, and the audio automatically adapts with scene changes, including environmental sound effects and character dialogues, filling in the "auditory" puzzle piece.
Genie 4 emphasizes the highest fidelity, while Genie 3 continues to be available in the Project Genie model selector for longer exploration and panoramic scenes.
With the launch of Genie 4, Google also introduced Project Genie Agent. Previously, creating a world required carefully organizing prompts; now it has become a conversation. You describe the world you want in natural language, and the agent will coordinate Google's multimodal model combinations to generate the initial visuals to establish the world's visual style, accompanied by custom background music.
More interestingly, the Gemini models can observe the screen content in real-time as you explore the world, actively popping out suggestions for "where to go next" or automatically triggering events within the world based on your journey.
You can also remix existing worlds in the gallery, adjusting the underlying structure, soundtrack, or rules, turning it into your own version, and then publish it in the community gallery for others to explore and create derivative works.
Currently, the new experience of Project Genie is open to eligible Google AI Ultra subscribers, while Google AI Pro members can join the waiting list on the same page.

免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。