Google officially launched Neon, a new model in the Gemini 4 series, on Wednesday. Evaluations by independent testing firm Artificial Analysis show that Neon scored 61 points on the Intelligence Index, ranking third globally, while its per-task cost fell about 17% compared with the previous generation Carbon and completion time shrank by roughly 40%.
On the same day, Google also introduced Stitch Loop, a long-running product agent, opened the Gemini App to low-memory Android devices, and officially began rolling out the world model Genie 4 to some users.
Neon's per-task cost is just $6.44, with 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 targets complex software engineering, knowledge work and scientific research tasks, with a focus on improving the model's end-to-end execution in complex workflows while balancing speed and cost.
Results released by Artificial Analysis show that Neon scored 61 points on the Artificial Analysis Intelligence Index at the highest xhigh thinking level, one point higher than the high level of Gemini 4 Carbon, ranking third globally behind only GPT-6.2 Astra and Claude Fable 5.2. Neon showed modest gains in agentic knowledge work, visual reasoning and coding.
Efficiency gains are a highlight of this upgrade. Artificial Analysis estimates that Neon completes an intelligence index task at an average cost of $6.44, about 17% below Carbon's $7.76 under standard pricing, and takes about 21 minutes on average, roughly 40% shorter than Carbon's about 36 minutes.
Notably, Neon actually uses about 15% more output tokens than Carbon when completing tasks. Artificial Analysis believes the cost decline mainly comes from Neon's lower token unit price, while the shorter time results from faster output speed: Neon outputs about 110 tokens per second, nearly double Carbon's roughly 57 tokens.
Neon also stands out in coding agent tests. Measured by Artificial Analysis in the Antigravity CLI environment, Neon scored 59 points on the Coding Agent Index, 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 cost to complete tasks is about $6.86, lower than Claude Opus 5.5's $12.61 and Claude Fable 5.2's $15.47.
Neon also made progress in visual reasoning and deliverable presentation. In the AA-Briefcase test, Neon's overall score reached 1,700 Elo, up 89 points from Carbon's 1,611; its presentation quality score reached 1,771 Elo, 45 points higher than the previous leader. Still, its overall score remains below models such as Claude Fable 5.2 and Claude Sonnet 5.5.
Ranked first in coding agents, matching Claude on complex code tasks
Google positions Neon as a model balancing frontier intelligence with execution efficiency, emphasizing its ability to handle multi-file debugging, cross-codebase refactoring, and long-horizon tasks requiring successive tool calls.
Neon supports a 1 million token context window and can integrate large projects, spreadsheets, charts, PDFs, audio and video in a single task. Google says the model offers four adjustable thinking levels — low, medium, high and xhigh — letting users weigh response speed, reasoning depth and token overhead according to task complexity.
Google also highlighted Neon's progress in multi-step instruction following and reducing model fabrication. The company says Neon's fabrication rate in internal Antigravity tests hit a new low for Gemini models; in several of Google's enterprise, financial and scientific task evaluations, Neon also achieved top results, including legal research, tax agents, automation tasks, and scientific and mathematical reasoning.
According to Google's official blog, Neon ranks 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 level. In the SWE-sweep test, Neon scored 12.8% in the mini-SWE test environment and reached 39.0% in the Google Antigravity environment.
Google also demonstrated a /swarm task in which Neon ran for up to 12 hours in Antigravity: the model used archival photos to reconstruct Alan Turing's study at King's College, Cambridge in 1936 and built a working universal Turing machine in a 3D scene. Google presented this as a showcase of the model's ability to handle long-horizon, multi-step complex tasks.
API pricing below Carbon's standard rate, cybersecurity capabilities further strengthened
Neon's API pricing is $3 per million input tokens and $15 per million output tokens; cached input tokens cost $0.15 per million, a 95% discount.
This price is higher than Carbon's current limited-time promotional rate of $2 per million input tokens and $10 per million output tokens, but lower than Carbon's standard rates of $4 and $20, a 25% discount to standard pricing. Carbon's limited-time promotion ends on November 30.
Google also launched a cybersecurity version of Neon, focused on vulnerability discovery, risk validation and code remediation. Google says Neon reached 76% on the hard tier of AutoBaxBench and 51.26% on WebExploitBench v1.2; it scored 90.5% on the SEC-bench Pro vulnerability detection test and reached 70.7% on the CWE-bench v2 root-cause remediation test.
Google says Neon outperforms Carbon on cybersecurity tasks and in some tests costs about half as much as large frontier models. The Neon cybersecurity version will be opened to vetted cybersecurity defenders through the Fairwind program in the coming days.
On safety, Google says Neon continues Carbon's protections against severe abuse and prompt injection. In the Gray Swan indirect prompt injection test, Neon at its highest thinking level withstood all attacks in the test, with a 0% attack success rate. However, this result represents only a specific benchmark and does not mean the model can resist all prompt injection attacks in real-world environments.
Stitch Loop lets agents continuously monitor and optimize products
Beyond the base model, Google also launched Stitch Loop, a new experimental product from Google Labs, on Wednesday. It is a long-running product agent that can connect to enterprise codebases and real-time product metrics, continuously identify product issues, propose improvements, and — after user confirmation — write code and create GitHub Pull Requests.
Stitch Loop's workflow includes discovering improvement opportunities, proposing solutions, generating code and continuously tracking results. It can connect to GitHub as well as product data such as logs, user behavior and feedback, and obtain context through more than 180 integrations, custom MCP servers or local CLI uploads.
Google says Stitch Loop has built-in recurring tasks such as cybersecurity, latency and accessibility optimization, and also allows teams to set their own goals, such as improving checkout conversion. It can write code and submit Pull Requests in an isolated cloud sandbox, but will not automatically merge before user review.
Among these, the cybersecurity loop combines Gemini 4 Carbon's cybersecurity capabilities with Google DeepMind's autonomous agent CodeMender to scan codebases for vulnerabilities and propose fixes. Stitch Loop is currently opening a waitlist, with eligible Google AI Ultra users getting priority access.
Gemini App opens to low-memory Android devices, Genie 4 begins rolling out to users
Google also announced on Wednesday that the Gemini App now supports eligible low-memory Android Go devices. Google says it has redesigned the app so the core Gemini experience runs smoothly on lower-memory hardware, and expects this to give hundreds of millions of users worldwide access.
Newly supported features include Gemini Live voice conversations, recognizing notes and tables via camera for learning or translation, and generating images with Nano Banana. Users can download the app from Google Play, and some new devices from partners may also come preinstalled with Gemini in the future.
In addition, Google officially opened the world model Genie 4 to eligible Google AI Ultra subscribers and launched an upgraded Project Genie experience. Through natural language instructions, users can change weather, add characters or adjust scenes in real time within generated interactive worlds, and also experience native audio generated as the environment changes.
Project Genie also introduces a creative agent: users simply 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 made by others. Google AI Pro users can currently join a waitlist.
From Neon improving the execution efficiency of complex tasks, to Stitch Loop letting agents continuously participate in product iteration, to Genie 4 opening up interactive virtual worlds, Google's series of announcements on Wednesday shows its AI product strategy is extending from model capabilities themselves to a broader range of application scenarios. For Google, the key going forward is not only whether model benchmark scores can continue to rise, but also whether these tools can truly enter the daily workflows of developers, enterprises and ordinary users.
Comments