The development

What happened

Google has begun rolling out Gemini 4 Neon, an AI model for complex coding, analytical and enterprise work. The company positions it between Carbon's long-horizon role and Hydrogen's everyday execution, with a one-million-token context window and several thinking levels. Access is split across developer, enterprise and paid consumer products; the cyber variant has a separate phased rollout to eligible defenders.

Key facts

  • Google advertises $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. These are token prices, not a guarantee about the cost of completing a task.

  • Developer access begins through the Gemini API paid tier in Google AI Studio and Android Studio. Pro and Ultra subscribers can use Neon in the Gemini app and Antigravity; enterprise access includes Gemini for business.

  • Neon has four thinking levels. Google reports an Artificial Analysis Intelligence Index score of 61 at xhigh, distinct from its default high setting; TideReach has not independently reproduced the result.

  • Google reports 12.8% on SWE-sweep with the mini-SWE harness and 39.0% in Antigravity. Those harness-specific results should not be read as one model-only success rate across arbitrary repositories.

  • Gemini 4 Neon for cyber is scheduled for eligible trusted Fairwind Program defenders over the coming days. That is separate from the general model rollout and does not establish unrestricted cyber access.

The context

Why it matters

Neon adds a middle option for teams choosing among Google's reasoning models. Its practical value depends on the workload, thinking setting and surrounding tools: the launch's own harness comparison shows why a model score is not a complete agent-system result. Compare end-to-end accuracy, retries, latency and token spend on representative tasks rather than assuming advertised per-token prices or selected benchmark savings determine your costs. Google's safety and fabrication claims are not a guarantee of reliable or injection-proof execution.