AI Trends

Google Releases Gemini 4 Neon: Ranks Third Globally in Intelligence Index, With Task Costs Down About 17%

WallstreetcnUpdated Nov 19 09:45

AI summary

Google has released Gemini 4 Neon, which scored 61 on Artificial Analysis’s Intelligence Index, ranking third globally. Its average cost per task is $6.44, about 17% lower than Carbon, while average completion time has fallen from around 36 minutes to 21 minutes. Neon supports a 1 million-token context window and four thinking levels. API pricing is $3 per million input tokens and $15 per million output tokens. On the same day, Google also launched Stitch Loop, a long-horizon product agent, and made Genie 4 available to select users.

Why read this

SelectedAI score 86

The material combines third-party benchmarks with task cost and completion-time data, enabling comparisons of Gemini 4 Neon’s capabilities and execution efficiency relative to its predecessor.

Topics

Tags

Body · AI translation

Google officially unveiled Neon, a new model in the Gemini 4 series, on Wednesday. Evaluations by independent benchmarking organization Artificial Analysis show that Neon scored 61 on its Intelligence Index, ranking third globally, while its per-task cost fell by about 17% from its predecessor Carbon and completion time decreased by about 40%.

On the same day, Google also launched Stitch Loop, an agent for long-running product tasks, made the Gemini App available on low-memory Android devices, and officially opened access to the Genie 4 world model to select users.

Per-task cost of just $6.44, with completion time cut to 21 minutes

Koray Kavukcuoglu, senior vice president and chief AI architect at Google DeepMind, said in an official blog post that Gemini 4 Neon is designed for complex software engineering, knowledge work, and scientific research tasks. It focuses on improving end-to-end execution across complex workflows while balancing speed and cost.

Artificial Analysis' evaluation results show that Neon scored 61 on the Artificial Analysis Intelligence Index at the highest xhigh reasoning level, one point above Gemini 4 Carbon at high, ranking third globally behind only GPT-6.2 Astra and Claude Fable 5.2. Neon made modest gains in agentic knowledge work, visual reasoning, programming, and other areas.

Efficiency gains are a major highlight of this upgrade. Artificial Analysis estimates that Neon's average cost to complete an Intelligence Index task is $6.44, about 17% below Carbon's $7.76 at standard pricing. Its average completion time is about 21 minutes, roughly 40% shorter than Carbon's approximately 36 minutes.

Notably, Neon uses about 15% more output Tokens than Carbon when completing tasks. Artificial Analysis believes the cost reduction primarily comes from Neon's lower per-Token price, while the shorter completion time results from faster output: Neon generates about 110 Tokens per second, nearly twice Carbon's approximately 57 Tokens per second.

Neon also performed strongly in coding-agent tests. In the Antigravity CLI environment, Artificial Analysis measured a Coding Agent Index score of 59 for Neon, three points higher than Carbon and roughly on par with Claude Code running Claude Fable 5.2 and Claude Opus 5.5. However, Neon's average task cost was about $6.86, below the $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2.

Neon also improved in visual reasoning and the presentation of results. In the AA-Briefcase benchmark, Neon achieved an overall score of 1700 Elo, up 89 points from Carbon's 1611. Its presentation-quality score reached 1771 Elo, 45 points above the previous leader. Its overall score, however, remained below models including Claude Fable 5.2 and Claude Sonnet 5.5.

Ranks first among coding agents, matching Claude on complex coding tasks

Google positions Neon as a model that combines frontier intelligence with execution efficiency, emphasizing its ability to handle multi-file debugging, cross-codebase refactoring, and long-running tasks requiring continuous tool use.

Neon supports a 1 million-Token context window, allowing it to integrate information from large projects, spreadsheets, charts, PDFs, audio, video, and other sources within a single task. Google says the model offers four adjustable reasoning levels—low, medium, high, and xhigh—allowing users to balance response speed, reasoning depth, and Token usage according to task complexity.

Google also highlighted Neon's progress in following multi-step instructions and reducing model fabrications. The company said Neon achieved the lowest fabrication rate of any Gemini model in internal Antigravity testing. Neon also ranked near the top in several Google evaluations covering enterprise, financial, and scientific tasks, including legal research, tax representation, automation tasks, and scientific and mathematical reasoning.

Google's official blog disclosed that Neon ranked first on the Artificial Analysis Coding Agent Index with a score of 58.6. It scored 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting. In the SWE-sweep benchmark, Neon scored 12.8% in the mini-SWE test environment and 39.0% in the Google Antigravity environment.

Google also demonstrated Neon running a /swarm task in Antigravity for as long as 12 hours: the model used archival photographs to reconstruct Alan Turing’s study at King’s College, Cambridge, in 1936, and built a functioning universal Turing machine within the 3D scene. Google presented this as a demonstration of the model’s ability to handle long-horizon, complex, multi-step tasks.

API pricing below Carbon’s standard rates as cybersecurity capabilities continue to improve

Neon’s API is priced at $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.15 per million, a 95% discount.

This is higher than Carbon’s current limited-time promotional price of $2 per million input tokens and $10 per million output tokens, but 25% below Carbon’s standard rates of $4 and $20. Carbon’s limited-time promotion will end on November 30.

Google also introduced a cybersecurity version of Neon focused on vulnerability discovery, risk validation, and code remediation. According to Google, Neon achieved 76% on the difficult tier of AutoBaxBench and 51.26% on WebExploitBench v1.2. It scored 90.5% on the SEC-bench Pro vulnerability-detection benchmark and 70.7% on the CWE-bench v2 root-cause remediation benchmark.

Google said Neon outperforms Carbon on cybersecurity tasks and costs about half as much as large frontier models on some benchmarks. The cybersecurity version of Neon will become available to vetted cyber defenders through the Fairwind program in the coming days.

On safety, Google said Neon maintains Carbon’s protections against severe misuse and prompt injection. In Gray Swan’s indirect prompt-injection test, Neon resisted every attack at its highest thinking setting, for an attack success rate of 0%. However, this result reflects only a specific benchmark and does not mean the model can withstand every prompt-injection attack encountered in real-world environments.

Stitch Loop enables agents to continuously monitor and optimize products

In addition to the foundation model, Google on Wednesday released Stitch Loop, a new experimental product from Google Labs. It is a long-running product agent that can connect to a company’s codebase and real-time product metrics, continuously identify product issues and propose improvements, and, after receiving user approval, write code and create GitHub pull requests.

Stitch Loop’s workflow includes identifying opportunities for improvement, proposing solutions, generating code, and continuously tracking results. It can connect to GitHub and product data such as logs, user behavior, and feedback, while also obtaining context through more than 180 integrations, custom MCP servers, or uploads from a local CLI.

Google said Stitch Loop includes built-in loops for tasks such as cybersecurity, latency, and accessibility optimization, while also allowing teams to set their own goals, such as increasing checkout conversion rates. It can write code and submit pull requests in an isolated cloud sandbox, but it will not merge them automatically before user review.

The cybersecurity loop combines Gemini 4 Carbon’s cybersecurity capabilities with CodeMender, an autonomous agent from Google DeepMind, to scan codebases for vulnerabilities and propose fixes. Stitch Loop currently has a waitlist, with eligible Google AI Ultra users receiving priority access.

Gemini App expands to low-memory Android devices as Genie 4 begins rolling out to users

Google also announced Wednesday that the Gemini App now supports eligible low-memory Android Go devices. Google said it redesigned the app so the core Gemini experience can run smoothly on hardware with less memory, potentially giving hundreds of millions of users worldwide access to it.

Newly supported features include Gemini Live voice conversations, using the camera to recognize notes and tables for learning or translation, and generating images with Nano Banana. Users can download the app from Google Play, and new devices released by some partners in the future may also come with Gemini preinstalled.

In addition, Google officially made the Genie 4 world model available to eligible Google AI Ultra subscribers and introduced an upgraded Project Genie experience. Using natural-language instructions, users can change the weather, add characters, or adjust scenes in real time within generated interactive worlds. They can also experience native audio generated dynamically as the environment changes.

Project Genie is also introducing a creation agent. Users need only describe an idea, and the agent can coordinate Google’s multimodal models to generate an initial visual scene and background music. A newly launched community gallery allows users to publish, explore, and remix virtual worlds created by others. Google AI Pro users can currently join the waitlist.

From Neon improving the efficiency of complex-task execution, to Stitch Loop enabling agents to participate continuously in product iteration, to Genie 4 opening up interactive virtual worlds, Google’s series of announcements on Wednesday shows that its AI product strategy is expanding beyond model capabilities themselves into a broader range of applications. For Google, the next key question is not only whether its models can continue improving their benchmark scores, but also whether these tools can truly become part of the everyday workflows of developers, businesses, and ordinary users. Risk Warning and Disclaimer

Markets involve risks, and investments should be made with caution. This article does not constitute personalized investment advice, nor does it take into account the specific investment objectives, financial circumstances, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are appropriate for their particular circumstances. Anyone investing on this basis does so at their own risk.

Source

Wallstreetcn