
Google has added another member to its Gemini 4 lineup. Gemini 4 Neon is now available to Pro and Ultra subscribers of the Gemini app, to developers and to businesses, positioned as a model that balances depth and efficiency for demanding but frequent workloads.
Koray Kavukcuoglu, senior vice president at Google DeepMind and the company’s chief AI architect, called Neon a frontier-level model, meaning it is meant to stand alongside the most capable systems on the market when handling coding, office work and scientific tasks.
- Key Point: Neon costs developers $3 per million input tokens and $15 per million output tokens, with a context window of one million tokens.
- Key Point: Cached tokens are priced at $0.15 per million, a 95% discount, according to Google.
- Key Point: Google says Neon costs 32% less per task than Carbon at its default reasoning level, and 50% to 80% less than leading rival models.
- Key Point: A cybersecurity-tuned variant is slated to reach trusted defenders in the Fairwind program within days and will not be publicly released.
Where Neon sits in the Gemini 4 family
Neon fills the space between two siblings. Carbon is reserved for the hardest, longest-running projects, while Hydrogen handles quick everyday requests. Neon targets the middle ground: complex work that recurs often enough that cost and speed matter as much as raw capability.
Developers can try it through the paid tier of the Gemini API, in Google AI Studio and in Android Studio. Pro and Ultra subscribers also get access inside Google Antigravity, the company’s agentic coding platform, where AI assistants carry out tasks on their own. For enterprises, the model is reachable through Gemini for business, either by API or from the model menu.
Benchmark claims and safety posture
Google says Neon ranks third overall with 61 points on the Artificial Analysis Intelligence Index, a comparison service, and second on the Vals index at 70.9%. Those evaluations come from third parties, though Google chose which ones to present. On coding, the company says Neon tops the Artificial Analysis coding agent index at 58.6 points, and that it scored 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. These are comparative tests rather than measures of real-world use.
The company also says Neon carries over Carbon’s defenses against misuse and prompt-injection attacks, in which malicious instructions are hidden in web pages, emails or documents the model reads. It reports that Neon declined only 1.98% of tasks in the Terminal-Bench 4.0 test.
A security-focused edition and what comes next
A version tuned for cybersecurity is in preparation, aimed at finding, validating and fixing vulnerabilities in code. According to the announcement, it will reach trusted defenders in the Fairwind program “in the coming days” and will not be open to the general public.
The pricing and benchmark framing point to a clear commercial strategy: Google is trying to make the case that a mid-tier model can absorb high-volume agentic and coding work without the per-task expense of its top-end systems. Whether enterprises accept that trade-off will depend on independent testing, since the numbers presented so far are ones Google selected. The limited Fairwind rollout will also be watched closely as an early signal of how the company intends to keep security-sensitive capabilities out of unrestricted hands.
