Google’s Gemini 4 Neon Targets the Cost-Performance Sweet Spot

Google has expanded its Gemini 4 family with the launch of Gemini 4 Neon, a model that prioritizes cost-efficiency and speed over raw benchmark supremacy. Announced on November 18, 2026, by Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect, Neon slots between the high-end Gemini 4 Carbon and the lightweight Gemini 4 Hydrogen, offering what Google calls “frontier-level intelligence” for complex, high-frequency workflows.

  • Positioned for balance: Neon sits between Carbon (long-horizon, heavy tasks) and Hydrogen (fast everyday execution), targeting enterprise, coding, and scientific workloads.
  • Aggressive pricing: $3 per million input tokens, $15 per million output tokens, and $0.15 per million cached input tokens—a 95% discount for reused context.
  • Strong agentic performance: Ranks #3 overall on the Artificial Analysis Intelligence Index (61 points) and #2 on the Vals Index (70.9%), with top scores in coding and knowledge-work benchmarks.

Where Neon Fits in the Gemini 4 Lineup

Google’s decision to split Gemini 4 into three distinct models reflects a broader industry shift: not every task requires the most expensive engine. Carbon handles the hardest, longest-running projects; Hydrogen manages quick, everyday executions; and Neon bridges the gap, delivering frontier-level reasoning for high-frequency tasks at a lower cost per task. According to Google, Neon is adapted from Gemini 4 Carbon, inheriting its architecture and safety mitigations while being tuned for efficiency.

The model maintains a 1-million-token context window and introduces four “thinking” levels, including a default high mode and an xhigh mode for deeper multi-step reasoning. This granularity lets teams dial intelligence and token spend to match each workload—a practical refinement rather than an architectural leap.

Benchmarks: Strong in Agentic Tasks, Not the Overall Leader

Google’s published results show Neon excelling in agentic and coding scenarios. It ranks #1 on the Artificial Analysis Coding Agent Index (58.6), scores 70.5% on Terminal-Bench Science, 71.3% on HLE-Diamond, and 84.0% on RiemannBench. In the Vending-Bench 2 simulation of autonomous business management, Neon generated $16,843 in simulated revenue—the highest in the table. It also leads on Legal Research Bench and Tax Agent Bench, and ranks #2 on Arena’s new Agent Arena Index with a score of 1336 across 4,200+ real-world sessions.

However, Google openly admits Neon is not the top model in raw intelligence. On the Artificial Analysis Intelligence Index, it scores 61 at xhigh, placing third overall—”just behind larger, slower models.” Notably, Google’s announcement does not name competitors, referring only to “competing models” or “larger, slower models.” The subtext is clear: Neon is nearly as good as the best, but significantly cheaper.

“Neon ranks ahead of comparable models and #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, landing just behind larger, slower models.” — Google’s official announcement

On cost, Google claims Neon is 32% cheaper per task than Carbon at its default high setting, and 36% cheaper than rival models at their highest settings. Against top competing frontier models, the company cites 50% to 80% lower cost per task on third-party benchmarks. These figures position Neon as a cost-effective alternative for high-volume agentic workflows.

Security, Availability, and a Cyber-Focused Variant

Google is also emphasizing safety and reliability. Neon rejected only 1.98% of tasks in Terminal-Bench 4.0, recorded a 0.09% “misaligned action” rate, and achieved 99.2% on an agent honesty test. These internal metrics, while not independently verified, suggest a focus on production-grade deployment. The model also produces more secure code than Carbon across easy, medium, and hard challenges on AutoBaxBench, according to Google.

A specialized “Gemini 4 Neon for Cyber” variant is slated to arrive in the coming days for trusted defenders within Google’s Fairwind program. It targets defensive cybersecurity tasks such as secure code generation, vulnerability detection, penetration testing, and patching. Early results show 90.5% on SEC-bench Pro and 51.3% on WebExploitBench v1.2.

Neon is available immediately across the Gemini API (paid tier in Google AI Studio and Android Studio), Google Antigravity for AI Pro and Ultra subscribers, Gemini Enterprise for businesses, and the model selector in the Gemini app for Pro and Ultra users. The broad, same-day rollout follows Google’s pattern of pushing new models across its entire portfolio at once.

With Neon, Google is betting that most real-world work lives in the middle ground—not at the frontier. By offering a model that is “almost as good and much cheaper,” the company aims to win on cost per task while conceding the absolute performance crown, for now. The strategy is defensible, but it also acknowledges that the true leader remains “larger and slower”—and not made by Google.

Sources & References