Google Introduces Gemini 4 Neon
Google introduced Gemini 4 Neon on November 18, 2026. The post, by Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect at Google, prices the model at $3 per million input tokens and $15 per million output tokens. Cached input is $0.15 per million tokens, which Google calls a 95 percent discount. The context window is 1 million tokens. The scores below are the ones Google states. Ultrathink did not rerun them.
Where Google says it is available
Google says Neon is rolling out the same day in the Gemini API in Google AI Studio, in Gemini for business through the API and the model dropdown, for paid Google AI Pro and Ultra subscribers in Google Antigravity, and for Pro and Ultra subscribers in the Gemini app. A later section says developers can build with the Gemini API paid tier in Google AI Studio and Android Studio, and that Pro and Ultra subscribers can try Neon in AI Studio. The post does not say the model is on the free Gemini tier.
Google describes a separate variant, Gemini 4 Neon for cyber, rolling out to trusted defenders in the Fairwind program in the coming days. That is not the general rollout. Ultrathink's earlier article covered Gemini 4 Argon, a different Fairwind model. This post is Neon.
Google says Neon has four thinking levels and that high is the default. The post names xhigh as the deeper setting for harder engineering, analytical, and science work. It does not name the other two levels in the sections cited here. It also does not state a parameter count, a training cutoff, or a knowledge cutoff.
Scores Google cites
- Artificial Analysis Intelligence Index: 61 at xhigh. Google says that is ahead of comparable models, third overall, and behind larger, slower models.
- Vals Index: second, at 70.9 percent. Google also says Neon is on the Pareto frontier of that index.
- Arena Agent Arena Index: 1336, second, across more than 4,200 sessions. Google also says Neon is second on steerability in that index.
- Coding, as Google states it: first on the Artificial Analysis Coding Agent Index at 58.6, 65.2 percent on Terminal-Bench 4.1, and 46.2 percent on FrontierCode 2.0 at the default high setting. On SWE-sweep, Google says 12.8 percent on the mini-SWE harness and 39.0 percent in Google Antigravity.
- Other scores in the post: 73.5 percent on Chartography, 75.8 percent on Agents' Last Exam v1.1, first on Zapier's AutomationBench at 56.3 percent, and first on Andon Labs' Vending-Bench 2 at $16,843. Google says Neon is first on both the Legal Research Bench and the Tax Agent Bench, and does not print those two scores. On science tasks, Google says Neon is best in class on Terminal-Bench Science at 70.5 percent, RiemannBench at 84.0 percent, and HLE-Diamond at 71.3 percent.
- Cost, in Google's comparisons: at the default high setting, a 32 percent lower cost per task than Gemini 4 Carbon on the Artificial Analysis Intelligence Index comparison Google describes, and about 50 percent to 80 percent lower cost per task than the top competing frontier models on the Vals index and Agent Arena. At xhigh, Google says the cost per task is still 36 percent or more below competing models at their highest setting.
The cyber variant, and the safety figures
For Gemini 4 Neon for cyber, Google says the model produces more secure code than Carbon on easy, medium, and hard AutoBaxBench challenges, and calls the hard tier state of the art at 76 percent. Google says WebExploitBench v1.2 is 51.26 percent at half the cost of competing large models, SEC-bench Pro is 90.5 percent, and Collinear's CWE-bench v2 is 70.7 percent, a 3.4 percent lift over Carbon. On ExploitGym, Google says the result is frontier-level and uses half the tokens of Carbon. Access to that cyber variant is the Fairwind rollout, not the general API rollout.
On safety, Google says real-time safeguards are comparable to its Carbon protections, and that Neon refused 1.98 percent of tasks on Terminal-Bench 4.0 and 0 percent on Terminal-Bench Science. Google says cheating on Terminal-Bench 4.0 was 0.0 percent, shortcuts on its coding cheating evaluation were 2.45 percent, the misaligned-action rate on its agentic safety suite was 0.09 percent, and the pass rate on its agent honesty evaluation was 99.2 percent. On Gray Swan's indirect prompt-injection benchmark, Google says Neon resisted every attack in that set at its highest thinking level with real-time safeguards, a 0.0 percent attack success rate it says other Gemini models have also reached. Google also says this is its lowest fabrication rate yet in internal Antigravity testing, and does not print that rate.
Google points readers to a Gemini 4 Neon model card for more detail. The figures in this article come from the blog post.
This article was ultrathought.