Google released Gemini 4 Neon on November 18, 2026, positioning it as the pragmatic choice for developers who want Carbon-level intelligence without the Carbon price tag. As the fourth model in the Gemini 4 family in just seven weeks, Neon costs $3 per million input tokens and $15 per million output tokens. It scores 60 on the Artificial Analysis Intelligence Index at default high effort, matching Gemini 4 Carbon but at roughly half the cost per task ($3.74 vs. $7.76).

The Economics of Verbosity

The sticker price is deceptive because Neon is inherently verbose. While the input/output rates look competitive, the model’s tendency to write extensive reasoning chains drives up actual usage. At the 'high' setting, Neon writes significantly more tokens than competitors like GPT-6.2 Sol. The 'xhigh' setting only adds 1.8 points to the intelligence score but costs an additional $2.70 per task and nearly seven extra minutes of processing time. For most workflows, the 'high' setting is the only rational choice; 'xhigh' is essentially paying a premium for marginal gains that chat interfaces can’t even access due to default restrictions.

Coding Benchmarks and Real-World Tests

In independent coding tests, Neon demonstrated strong efficiency. On MindStudio’s KingBench 4 test, it ranked second out of 18 models, trailing only Claude Fable 5.2. A full six-app run cost just $20.36 on Neon, compared to $222 for Fable 5.2 and $234 for Carbon. Neon produced 70–95% correct code on the first try across these builds. However, it lost to Opus 5.5 by a narrow margin on Terminal-Bench 4.1 and FrontierCode 2.0, suggesting that while Neon is efficient, it may lack the nuanced edge of Anthropic’s top models for complex, reviewer-level code changes.

Security Flaws and Context Limitations

Developers must be wary of Neon’s approach to security and context management. In one test, Neon generated clean Go code for a Docker alternative that failed to prevent hostile container images from escaping via seven symlink paths—a flaw Opus 5.5 and GPT-6.2 models avoided. Furthermore, Neon now shares a single 1,048,576-token budget for input, thinking, tool calls, and output. Unlike Gemini 3, where these limits were separate, a long reasoning run in Neon can exhaust the entire context window, breaking complex agent workflows if not carefully monitored.

Key Takeaways

  • Neon matches Gemini 4 Carbon’s intelligence score (60) at half the cost per task ($3.74 vs. $7.76).
  • The model is prone to token bloat; the 'xhigh' reasoning setting offers minimal returns for a significant cost increase.
  • Security testing revealed vulnerabilities in container isolation that top competitors like Opus 5.5 handled better.
  • A unified token budget means heavy reasoning can starve the context window, requiring careful prompt engineering.
  • Day-one integrations include Cursor, GitHub Copilot, and LangChain, but no free API tier exists for Neon.

The Bottom Line

Neon is the new default for cost-conscious coding agents, but its security blind spots and token-hungry nature mean it’s not a drop-in replacement for high-stakes production environments yet.