Google has launched Gemini 4 Neon, a model it positions as frontier-level intelligence at higher speed and lower cost. It costs $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window.
Google says Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, which puts it third overall and just behind larger, slower models. Most figures in this article are benchmark claims that outside testers have yet to reproduce.
Pricing and where it is available
Cached input tokens cost $0.15 per million, a 95% discount. The model is rolling out in the Gemini API in Google AI Studio, in Gemini for business through the API and model dropdown, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra subscribers.

Benchmark claims for work tasks
Google pitches Neon at spreadsheets, PDFs, charts, audio and video, all within one context window. Its listed results:
- #2 on the Vals Index at 70.9%, and #1 on the Legal Research Bench and Tax Agent Bench.
- 73.5% on Chartography for reading dense professional charts, and 75.8% on Agents’ Last Exam v1.1, on par with leading models.
- #1 on Zapier’s AutomationBench (56.3%) and on Andon Labs’ Vending-Bench 2 ($16,843).
- #2 globally on Arena’s new Agent Arena Index with 1336 points across more than 4,200 real agent sessions.
For science and math, Google reports 70.5% on Terminal-Bench Science, 84.0% on RiemannBench and 71.3% on HLE-Diamond, calling Neon best in class on all three.

Coding and agent performance
Neon ranks first on the Artificial Analysis Coding Agent Index at 58.6. It scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. On SWE-sweep it gets 12.8% on the mini-SWE test setup, and 39.0% inside Google Antigravity, which Google says shows what the Antigravity setup adds.
The model is built for long agent sessions. It can run alone or act as an orchestrator for parallel Hydrogen subagents in Antigravity, and Google says it avoids redundant file reads, reports progress and handles context compaction in long sessions.

Cost per task
At the default high thinking level Neon costs 32% less per task than Carbon, Google’s other Gemini 4 model, according to third-party benchmarks. Google also says it costs roughly 50% to 80% less per task than top competing frontier models at that level, and at xhigh it costs 36% or more below rivals at their highest setting. Four thinking levels let teams trade intelligence against token spend.
A cyber-defense version
A separate Gemini 4 Neon for cyber covers finding, validating and patching vulnerabilities. Google reports:
- State of the art on SEC-bench Pro at 90.5%, at about half the cost per challenge.
- Frontier-level results on ExploitGym using half the tokens of Carbon.
- 70.7% on Collinear’s CWE-bench v2, a 3.4% uplift over Carbon.
- 76% on the hard tier of AutoBaxBench for secure code, and 51.26% on WebExploitBench v1.2 at half the cost of competing large models.
It arrives for trusted defenders in the coming days under the Fairwind Program.
Safety numbers
Google says Neon keeps Carbon-level protection against severe misuse and prompt injection while refusing less legitimate work: 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science. It reports 0.0% cheating on Terminal-Bench 4.0, a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on its honesty test. On Gray Swan’s prompt-injection benchmark, Neon with real-time safeguards at its top thinking level resisted every attack.
Built for deeper analysis in the Gemini app
Google also points to the Gemini app, where Neon is available to Pro and Ultra subscribers for work that needs sustained reasoning. The company’s announcement describes Neon as faster than larger Gemini 4 models, which is the main trade it is selling: slightly lower peak scores for much lower cost per task.
What to check before switching
Most of these results are Google’s own or from third-party leaderboards, so teams should test Gemini 4 Neon on their own workloads before moving production traffic. The cost claims are the most useful to verify, since they depend on thinking level and task mix. Google’s model card has more detail.
Source: Google