Google Launches Gemini 4 Neon With 3x Faster Performance at Lower Cost
Google has introduced Gemini 4 Neon, a new artificial intelligence model designed to handle complex coding, research, business, and scientific tasks while offering faster performance and lower operating costs than larger AI models.
Google has introduced Gemini 4 Neon, a new artificial intelligence model designed to handle complex coding, research, business, and scientific tasks while offering faster performance and lower operating costs than larger AI models.
Article outline
- What happened
- The key numbers
- The details
- What comes next
- Background
- The bottom line
Key points
- While achieving 73.5% on Chartography and 75.8% on Agents' Last Exam v1.1, the model ranks first on the Legal Research Bench and Tax Agent Bench.
- Another demonstration indicated Neon converting a single image into a Google Sheet containing 100, 000 cells and more than 500, 000 formulas, without embedded images or Apps Script.
- It additionally leads Zapier's AutomationBench with 56.3% and Andon Labs' Vending-Bench 2, where it earned $16, 843 in a simulated business environment.
- In scientific reasoning, Google reports scores of 70.5% on Terminal-Bench Science, 84% on RiemannBench, and 71.3% on HLE-Diamond.
- It additionally achieves 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high thinking level.
Meanwhile, the firm states Neon combines advanced reasoning capabilities with efficient task execution, making it suitable for software development, enterprise workflows, and long-running AI agents.
According to early evaluations by Glean, Gemini 4 Neon matched or exceeded Gemini 4 Carbon in overall quality while operating three times faster and at 60% of the cost.
In practice, the model is rolling out through the Gemini app, Google AI Studio, Gemini API, Google Antigravity, and Google's business AI services. Claude Maker Anthropic Hits $2 Trillion, but There's a Catch. Performance and Benchmark Results.
Google notes Gemini 4 Neon ranks third overall on the Artificial Analysis Intelligence Index, scoring 61 at its highest thinking level.
Meanwhile, the model additionally ranks second on the Vals Index, with a score of 70.9%, and second on Arena's Agent Arena Index, scoring 1, 336 throughout more than 4, 200 real-world AI agent sessions.
On the new Agent Arena steerability evaluation, Neon ranks second for adapting to user corrections during ongoing tasks.
Google additionally reports robust performance in financial analysis, accounting, spreadsheet modeling, legal research, and tax-related tasks.
These results come from a combination of independent evaluations and benchmarks documented by Google. Improved Coding and AI Agents.
Gemini 4 Neon is designed for complex software development, including debugging, refactoring, working throughout substantial codebases, and coordinating multiple AI agents.
In practice, the model ranks first on the Artificial Analysis Coding Agent Index with a score of 58.6.
On SWE-sweep, Neon scores 12.8% using the mini-SWE harness, increasing to 39% when operating through Google Antigravity.
Google notes the model performs better during long-running tasks by reducing unnecessary file reads, maintaining context, and sharing progress updates.
In Antigravity, Neon can additionally coordinate multiple Gemini 4 Hydrogen subagents to complete larger projects.
Google demonstrated these capabilities through a number of projects, including a photorealistic 3D laboratory built using Blender, a WebGL2 skiing simulator based on real Mont Blanc terrain, and an interactive recreation of Alan Turing's 1936 study.
Google has additionally improved Neon's ability to identify and fix software security vulnerabilities.
Notably, the model achieved 76% on the hardest AutoBaxBench security challenges and 51.26% on WebExploitBench v1.2, with Google reporting lower costs than competing sizeable models.
It additionally scored 90.5% on SEC-bench Pro for identifying vulnerabilities and 70.7% on Collinear's CWE-bench v2 for patching them.
On ExploitGym, Neon demonstrated competitive performance in validating whether vulnerabilities could be exploited while using half the tokens required by Gemini 4 Carbon.
Google notes these improvements could assist security teams identify, assess, and repair vulnerabilities more efficiently.
Notably, a specialized version, Gemini 4 Neon for cyber, will become available to eligible security professionals through Google's Fairwind Program over the coming days.
Google Maps Can Now Plan Entire Trips and Find Appointments for You. Lower Pricing and Four Thinking Levels.
Google has priced Gemini 4 Neon at $3 per million input tokens and $15 per million output tokens.
Cached input tokens cost $0.15 per million, representing a 95% discount compared with standard input pricing.
In practice, the model additionally supports a 1 million-token context window, allowing it to process substantial amounts of information throughout documents, spreadsheets, charts, images, audio, and video.
Neon offers four thinking levels, allowing developers to adjust reasoning effort based on task complexity and cost requirements.
According to Google, its default high thinking level costs 32% less per task than Carbon on the Artificial Analysis Intelligence Index.
Meanwhile, the firm additionally claims Neon can cost approximately 50% to 80% less per task than leading competing models on certain evaluations.
At its highest thinking level, Neon delivers its strongest reasoning results while reportedly costing at least 36% less per task than competing models at their highest settings. Google Claims Better Reliability and Safety.
Google notes Gemini 4 Neon has achieved its lowest fabrication rate yet in internal Google Antigravity testing.
Meanwhile, the model additionally demonstrated improvements in instruction after, resistance to malicious prompts, and staying within user-defined boundaries.
In Google's internal evaluations, Neon achieved a 99.2% pass rate on agent honesty tests and a 0.09% misaligned-action rate.
It recorded no cheating on Terminal-Bench 4.0 and took shortcuts in 2.45% of tasks in Google's coding evaluation.
For context, the firm additionally documented reductions of 43% to 56% in incomplete assistance and 15% to 35% in actions beyond user instructions, depending on the reasoning setting.
On Gray Swan's prompt injection benchmark, Neon resisted all tested attacks when operating with its strongest safeguards and highest thinking level.
Taken together, the developments around google Launches Gemini 4 Neon With 3x Faster Performance at Lower Cost point to a situation that is still moving, and the coming days should bring more clarity.




