Near-top coding and agent quality at $3/$15, without Carbon's slowness
Gemini 4 Neon matched Carbon on Artificial Analysis and debuted second on Agent Arena.
Gemini Neon is the mid-tier Gemini 4. Within a day of launch, independent boards had it level with Carbon on intelligence and second on agent work, at $3.74 a task.
The deeper look
The numbers that matter are third-party. At high effort, Artificial Analysis has Neon matching Gemini Carbon's 60 on its Intelligence Index at $3.74 per task. Arena debuts it second on Agent Arena, with a $2.82 median task. Vals puts it second too, about 0.4 points behind Claude Fable 5.2 at under half the cost per task. Google's claim of #1 on the Coding Agent Index is the maker's word.
Human reports lean the same way. Hacker News commenters call it the first Google model that feels good for coding, in the bracket of Claude Opus 5.5. A blogger's early look reached a similar view.
Carbon is Ultra-only and reviewers found it slow, which makes Neon the one most people can actually try. It is not free, so Gemini Hydrogen is the cheaper test.
This is for builders who already pay for a top coding model. Run it on your own repo before making it the default.
How to start: pick Gemini 4 Neon in Google AI Studio on a paid key, or in Antigravity.
$ curl https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "google/gemini-4-neon", "messages": [{"role": "user", "content": "Hello"}]}'