- Google DeepMind released Gemini 4 Neon on November 18, 2026, with a 1 million-token context window.
- Pricing is set at $3 per million input tokens and $15 per million output tokens, with cached inputs at $0.15 per million.
- Neon ranks first on Artificial Analysis’s Coding Agent Index and matches Gemini 4 Carbon’s Intelligence Index score of 60.
- Real-world testing shows Neon’s headline cost advantage can collapse at high “thinking” settings, where it costs more per task than Carbon.
What Happened
Google DeepMind released Gemini 4 Neon on November 18, 2026, continuing the Gemini 4 family’s element-based naming after Carbon’s release two weeks after Argon. The model supports a 1 million-token context window and is priced at $3 per million input tokens and $15 per million output tokens, per Marktechpost (Nov. 18, 2026).
Why It Matters
Neon ranks first on Artificial Analysis’s Coding Agent Index, giving Google a marketing edge in the coding-model segment where Anthropic’s Claude Opus 5.5 and OpenAI’s models compete directly. Cached input tokens drop to $0.15 per million, a 95% discount designed to encourage high-volume agentic workloads.
Technical Details
Neon scores 60 on Google’s Intelligence Index benchmark, matching Carbon’s displayed score. However, independent cost analysis found that once Carbon’s November discount is factored in, the two models’ estimated benchmark costs are close at $3.74 and $3.88 per task, and Google’s headline 32% cost-per-task advantage for Neon narrows or reverses under real API testing with thinking tokens enabled.
Who’s Affected
Developers building coding agents and enterprises evaluating per-task inference costs are the most direct audience, since thinking-token volume — not list pricing — determines actual spend. Teams benchmarking against Claude Opus 5.5 now have a third frontier option in the comparison set.
What’s Next
Google has not disclosed a successor model name or timeline. Independent benchmark sites continue publishing comparative pricing and task-cost data as real-world usage accumulates.