Google released Gemini 4 Neon on Wednesday. It sits between the fast Gemini 4 Hydrogen and the flagship Gemini 4 Carbon. Google says it’s built for complex coding, multi-step agent workflows and analytical reasoning. That makes four Gemini 4 models in seven weeks.
Price and availability
Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15. Carbon is on a promotional $2/$10 until November 30 and goes to $4/$20 after that, which is also what Claude Opus 5.5 costs. The context window is one million tokens, and there are four effort levels from low to extra high.
In the Gemini app, Neon is live now for AI Pro and Ultra subscribers and replaces both Gemini 4 Argon and the old 3.1 Pro. Developers get it through the API, the Gemini Enterprise Agent Platform and Antigravity.
What the independent numbers say
Google’s own benchmarks put Neon ahead of Claude Opus 5.5 and GPT-6.2 Sol in several categories. The more useful numbers come from Artificial Analysis. On its Intelligence Index, Neon at the highest setting scores 61 and lands in third place. Ahead of it are GPT-6.2 Astra at 62 and Claude Fable 5.2 at 65. Carbon scores one point less than Neon. A task costs $6.44 with Neon against $7.76 with Carbon, and takes about 21 minutes instead of 36.
If you use Claude, the Coding Agent Index is the one to look at. Neon scores 59 there, running in the Antigravity CLI. Claude Code with Fable 5.2 scores 58, and so does Claude Code with Opus 5.5. The average task costs $6.86 with Neon, $12.61 with Opus 5.5 and $15.47 with Fable 5.2. On presentation quality, Neon now tops Artificial Analysis’s ranking, ahead of Claude Sonnet 5.5.
A tie at half the price
I wouldn’t read a one-point gap on the coding index as a lead. For agent work, Neon and the two Claude models are level. The price gap is real, though. The same task costs more than twice as much on Fable 5.2, and Google is shipping new models almost weekly. If your team mostly uses Claude Code for long agent runs, it’s worth testing Neon on your own codebase. Anthropic hasn’t released a new model since Fable 5.2 on November 9.