Key points
- Google DeepMind announced Gemini 4 Neon on Nov 18, 2026, placing it between its Carbon and Hydrogen models.
- Using its xhigh setting, Neon reached 61 on the Artificial Analysis Intelligence Index, ranking #3 overall, Google reports.
- Neon costs $3 for each million input tokens and $15 for each million output tokens, and accepts up to 1 million tokens.
- A cybersecurity version is rolling out to trusted defenders through Google's Fairwind Program.
Google DeepMind has announced Gemini 4 Neon, a new member of its Gemini 4 family of AI models. The company describes it as a middle option that balances depth and efficiency, aiming for frontier-level reasoning on coding, analysis and science tasks at higher speed and lower cost. Every performance figure below is as reported in Google's own announcement; some come from third-party benchmarks and others from Google's internal testing, and DexScience has not independently checked them.
Google places Neon between Carbon, which it pitches for the hardest long-running projects, and Hydrogen, which it pitches for quick everyday jobs. According to the company, Neon earned a score of 61 on the Artificial Analysis Intelligence Index when set to xhigh, its deepest reasoning mode. That put it #3 overall, behind larger and slower models. Google also says Neon ranked #1 on the Artificial Analysis Coding Agent Index, scoring 58.6. Thinking levels are settings that let users choose how much reasoning, and therefore how much compute, the model spends on a task. Neon has four of them.
Google charges $3 for every million tokens fed into the model and $15 for every million tokens it produces. Tokens are the small chunks of text that AI models read and write. The model can take in up to 1 million tokens at once. At its default high thinking level, Google says Neon costs 32% less per task than Carbon.
On safety, the company says Neon keeps protections against severe misuse and prompt injection at the same level as Carbon, while refusing only 1.98% of tasks on Terminal-Bench 4.0. A cybersecurity version will reach trusted defenders over the coming days through the Fairwind Program. The main model is rolling out now in the Gemini API in Google AI Studio, in Gemini for business, and in the Gemini app for Pro and Ultra subscribers.
Why it matters
Google is pitching Neon as a way to get close to its top models' performance on complex work while paying less per task.
What we don't know yet
All benchmark results, rankings and safety figures are as reported by Google; several are from third-party leaderboards, while others, including most safety figures, come from the company's internal testing. Google points readers to its Gemini 4 Neon model card for more detail.
Terms in this story
- Token
- A small chunk of text, such as a word or part of a word, that an AI model processes and that is used to measure usage and price.
- Thinking level
- A setting that controls how much reasoning an AI model does before it answers, trading speed and cost for depth.
- Prompt injection
- An attack in which malicious instructions, either typed directly or hidden in content such as a web page, email or document, try to override what an AI model has been told to do.
Sources and verification
DexScience articles are produced with AI assistance and an automated fact-check against the linked sources. Spotted an error? Tell us and we will correct it openly.


