Google launches Gemini 4 Neon for coding and analysis
Google has introduced Gemini 4 Neon, a model aimed at software engineering, knowledge tasks and science. The company says it combines strong benchmark results with lower cost per task and a 1 million-token context window.

Key takeaways
- Google introduced Gemini 4 Neon on 18 November 2026 for coding, analysis and science tasks.
- Google says the model has a 1 million-token context window and four thinking levels.
- Google reports top or near-top benchmark results in coding, enterprise analysis and science.
- Google says Gemini 4 Neon for cyber can help find, validate and patch vulnerabilities.
In this article
What Google announced
According to Google's Gemini blog and Google DeepMind, Google introduced Gemini 4 Neon on 18 November 2026. The company describes the model as a system for quick completion of multi-step tasks. It targets software engineering, knowledge tasks and science.
Google says Neon can work with spreadsheets, PDFs, charts, audio and video. It also says the model can turn mixed inputs into structured outputs within a context window of 1 million tokens.
Benchmark results and rankings

Google reports that Gemini 4 Neon scored 61 on the Artificial Analysis Intelligence Index at the xhigh setting. The company says this placed the model third overall and ahead of comparable models, though behind larger and slower systems.
The model also ranked second on the Vals Index, with a score of 70.9%, according to Google. Google says Neon reached first place on the Legal Research Bench and Tax Agent Bench. It also says Neon ranked second on Arena's Agent Arena Index. The reported score was 1336 across more than 4,200 agent sessions.
Google says Neon ranked second for steerability, meaning it adapts to user corrections during a task. It also reports first place on most Surge AI benchmarks. These cover finance, accounting, spreadsheet modelling and complex constraint following. Google says Neon also ranked first on Zapier's AutomationBench with 56.3% and on Andon Labs' Vending-Bench 2 with $16,843.
The supplied material does not include independent verification, so the results remain unconfirmed by outside sources.
Coding, science and enterprise tasks
For coding, Google says Neon ranked first on the Artificial Analysis Coding Agent Index with a score of 58.6. It reports scores of 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting.
Google says the model improved on SWE-sweep from 12.8% on the mini-SWE harness to 39.0% when used inside Google Antigravity. It says Neon can run in Antigravity and direct parallel Hydrogen subagents.
For science and analysis, Google lists scores of 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond. It also says Neon scored 73.5% on Chartography for reading dense professional charts and 75.8% on Agents' Last Exam v1.1.
Google provided several examples of Neon working in Antigravity. These included a 3D lab bench in Blender and a ski simulator. They also included a recreation of Alan Turing's study. Another example was a 100,000-cell spreadsheet generated from a single image.
Price and thinking levels
Google says Gemini 4 Neon costs $3 per million input tokens and $15 per million output tokens. Cached input tokens are priced at $0.15 per million, which Google describes as a 95% discount.
The model has four thinking levels. Google says teams can adjust these levels to balance intelligence and token spend.
At the default high level, Google claims Neon has a 32% lower cost per task than Gemini 4 Carbon. At xhigh, it says Neon costs more than 36% less than competing models at their highest settings.
Cyber defence and what it means
Google also presents a cyber-focused version, Gemini 4 Neon for cyber. It says this version produces more secure code than Gemini 4 Carbon on AutoBaxBench. It also says Neon reached 76% on the hard tier.
Google reports top results on WebExploitBench v1.2 at 51.26% and SEC-bench Pro at 90.5%. It says the cyber model can help with vulnerability finding, validation and patching.
For security teams, the announcement points to code hardening and penetration testing support. Readers should consult Google's official guidance before deployment.
The launch matters because Google is positioning Neon as a lower-cost model for demanding work. Its claimed strengths in coding, analysis and security could affect how teams choose models for production tasks.
Sources
- Introducing Gemini 4 NeonGoogle Gemini
- Introducing Gemini 4 NeonGoogle DeepMind
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



