AI Trends

Google Launches Gemini 4 Neon, Opens Access to Genie 4 and Project Genie Agent

Sina TechUpdated Nov 19 09:41

AI summary

Google has launched Gemini 4 Neon and opened access to the Genie 4 world model and Project Genie Agent. Neon is three times faster than Carbon and costs 60% as much, offering a 1 million-token context window and four levels of reasoning depth. Genie 4 supports mid-session prompting and native audio and is currently available to eligible Google AI Ultra subscribers.

Why read this

AI score 87

The material summarizes the capabilities, costs, and availability of the two models, providing a basis for comparing their suitability for complex workflows and interactive world generation.

Topics

Tags

Body · AI translation

New public testing promotional copy New public testing promotional copyend Add app-exclusive traffic redirect article_adlist[

Google has recently been pushing hard on foundation models, continuously releasing models ranging from the most intelligent to the most efficient, going all out in its bid to reclaim the throne.

Following Carbon, Argon, and Hydrogen, Gemini 4 Neon has arrived to fill the “mid-tier” gap between the most intelligent and most efficient models!

Demis Hassabis called the mid-tier Gemini 4 Neon model “Alchemy at its finest.”

In the Gemini 4 family, Carbon is the flagship, responsible for the most difficult long-horizon projects; Hydrogen targets frequent everyday scenarios, prioritizing response speed. Neon sits between the two, “providing frontier reasoning capabilities for complex coding, analytical reasoning, and production-grade enterprise workflows, while being faster and more economical.”

Neon is three times as fast as Carbon and costs just 60% as much. It is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens costing as little as $0.15. It has a 1-million-token context window and also supports four adjustable thinking-depth levels, allowing developers to flexibly balance speed and reasoning depth according to task complexity.

Gemini 4 Neon ran a 12-hour /swarm to reconstruct Alan Turing’s 1936 research from archival photographs—including a working universal Turing machine, the blueprint for modern computing.

Comprehensive benchmark results

Neon delivered strong results across a broad range of third-party evaluations. In general intelligence, it scored 61 on the Artificial Analysis Intelligence Index, ranking third and trailing only larger, slower models; it ranked second on the Vals Index at 70.9%, placing it on the Pareto frontier.

On the brand-new Agent Arena Index, Neon ranked second with 1,336 points based on more than 4,200 real-world agent sessions. It also placed second in steerability, meaning that when users change their instructions midway through a task, the model can adjust direction more cleanly.

It dominated the leaderboards even more extensively in enterprise and knowledge-work evaluations. It ranked first on both Legal Research Bench and Tax Agent Bench, and also placed first in most of the finance, accounting, spreadsheet modeling, and other categories covered by Surge AI. It ranked first on Zapier AutomationBench at 56.3%, and also took first place on Andon Labs’ Vending-Bench 2 with $16,843. The latter simulates long-term business operations and places heavy demands on long-horizon planning. Neon scored 73.5% on Chartography, a benchmark for reading dense charts, and 75.8% on Agents' Last Exam v1.1, matching the leading models.

In scientific and mathematical reasoning, Neon took first place on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%), achieving best-in-class results on all three.

A capability-by-capability breakdown of Neon

According to the official blog, Neon’s new capabilities can be grouped into several broad areas.

More accurate, more reliable coding. Neon is designed for real production environments and can handle multi-file debugging, refactoring, and precise edits. Google specifically highlighted its “surgical edits” capability, which enables targeted modifications to multi-file projects. The model has improved significantly at following multi-step instructions and recorded the lowest fabrication rate to date in internal Google Antigravity testing.

A 1-million-token window that handles multimodal inputs. Spreadsheets, PDFs, charts, audio, and video—Neon can process all these inputs within a single context window and produce structured deliverables. “Neon matches or even surpasses Carbon in overall quality, while being three times faster and 40% cheaper. It strikes a compelling balance between accuracy, detail, and speed.”

Cybersecurity. This is a distinctive element of the Neon release. Google has introduced Neon for Cyber, covering the entire chain from secure code generation and vulnerability discovery to vulnerability validation and patching.

According to the data, AutoBaxBench secure code generation reached SOTA at the difficult level (76%), outperforming Carbon across the board; WebExploitBench v1.2 penetration testing likewise achieved SOTA (51.26%) at only half the cost of competing large models; SEC-bench Pro vulnerability discovery reached SOTA (90.5%), with a per-challenge cost roughly half that of comparable models; and CWE-bench v2 vulnerability patching reached 70.7%, an improvement of 3.4% over Carbon. Neon for Cyber will become available to trusted security defenders through the Fairwind Program in the coming days.

Safety alignment with virtually “zero cheating.” Google has invested heavily in safety. Neon recorded a 0.0% cheating rate on Terminal-Bench 4.0, matching Carbon; only 2.45% of tasks in the coding-cheating evaluation involved shortcut-taking behavior, the lowest rate across the entire Gemini 4 family; and in the agent safety suite, it had a 0.09% misconduct rate and a 99.2% pass rate on honesty evaluations. On Gray Swan’s indirect prompt-injection benchmark, Neon achieved a 0.0% attack success rate at the highest thinking level. Compared with Carbon, Neon reduced “underserving” by 43%-56% and “overstepping” by 15%-35%. At the same time, it almost never refuses legitimate tasks, with a refusal rate of just 1.98% on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science.

Upgrades are also coming to consumers. In the Gemini app, Neon delivers higher-quality writing, can interpret uploaded spreadsheets, PDFs, photos, and other content more accurately, and can connect to Gmail and Docs to retrieve relevant information. The Tasks feature is now powered by Neon, making complex, multi-step tasks faster and more quota-efficient.

Gemini 4 Neon has begun rolling out. Developers can build applications in Google AI Studio and Android Studio through the paid tier of the Gemini API, while Google AI Pro and Ultra subscribers can try it in Google Antigravity. Enterprise users can access it through the Gemini for Business API or the model drop-down menu. On the consumer side, Pro and Ultra subscribers can switch directly to Neon in the Gemini app’s model selector. Neon for Cyber will be rolled out to trusted defenders through the Fairwind Program in the coming days.

World model Genie 4 finally launches

On October 28, Google DeepMind officially unveiled its latest-generation world model, Genie 4, calling it “the most realistic and interactive world model yet.”

After a wait of more than half a month, Google has finally announced that Genie 4 is officially available.

Genie 4 is Google’s most realistic and interactive world model to date. Its most important new capability is “mid-session prompting,” which allows you to alter the environment using natural language at any point while exploring a world. You can change the weather, introduce a character, or alter the direction of the story, with every change occurring in real time within the current scene.

Another capability making its debut is native audio. Worlds generated by Genie 4 come with sound, which automatically adapts as the scene changes, encompassing everything from ambient effects to character dialogue and completing the auditory piece of the puzzle.

Genie 4 focuses on maximum fidelity, while Genie 3 remains available in Project Genie’s model selector for longer explorations and panoramic scenes.

Alongside the launch of Genie 4, Google has also released Project Genie Agent. Previously, creating a world required carefully crafting a prompt; now it has become a conversation. You describe the world you want in natural language, and the agent orchestrates Google’s suite of multimodal models to generate initial visuals that establish the world’s visual style, accompanied by custom-tailored background music.

Even more interestingly, the Gemini model can observe the screen in real time as you explore the world, proactively offering suggestions about “where to go next” or automatically triggering in-world events based on your journey.

You can also remix existing worlds in the gallery, adjusting their underlying structure, soundtrack, or rules to turn them into your own versions, then publish them to the community gallery for others to explore and remix further.

The new Project Genie experience is currently available to eligible Google AI Ultra subscribers, while Google AI Pro members can join the waitlist on the same page. article_adlist[

]article_adlistarticle_adlist[

Please contact this official account for authorization to repost.

For submissions or media coverage requests: liyazhou@jiqizhixin.com ]article_adlist Statement: Exclusive Sina.com content; unauthorized reproduction is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end

Source

Sina Tech