Ask Gemini 4 Neon for a web page where the computer plays a full game of chess against itself, and it gets there. In a hands-on test published by DataCamp, editor Josef Waples gave Neon and Google’s bigger Gemini 4 Carbon the same prompt, three runs each. Neon built a working page every time, for $0.83 to $1.06 a run. Carbon finished once.

Carbon’s two failures had nothing to do with chess. It tried to write the whole program in one reply. That reply hit the 65,536 token output limit (tokens are small chunks of text) before the file was saved.

Neon wrote the page in several smaller edits. Three runs on one task is a small sample, and Waples says so himself.

What Google launched

Google announced Gemini 4 Neon on 18 November. It sits between Hydrogen, the fast everyday model, and Carbon, meant for the hardest long jobs. The Gemini 4 Neon model card says it is adapted from Carbon.

Google’s demos lean visual. In Antigravity, Google’s coding tool, Neon wrote a ski simulator on real Mont Blanc terrain with 10,200 lines of custom shaders. It also turned a single image into a 100,000-cell Google Sheet that relights the scene using more than 500,000 formulas. Those are Google’s own demos, and we found no outside attempt to rebuild them.

The model card says Neon reads text, images, audio and video. Input, thinking and output share one 1 million token window, according to Google’s developer notes for Neon.

Google’s numbers and the outside checks

Google’s model card compares Neon with Claude Opus 5.5 and GPT-6.2 Sol, not with Carbon. DataCamp counted Neon ahead on 17 of the 19 tests there, all scored by Google.

Artificial Analysis runs the same task set on every model, and its write-up on Gemini 4 Neon is cooler. At its deepest thinking setting, xhigh, Neon scores 61 on the firm’s Intelligence Index. That is one point ahead of Carbon and third overall, behind OpenAI’s GPT-6.2 Astra and Anthropic’s Claude Fable 5.2.

At high, Neon’s default setting, the model only matches Carbon. Our guide to what the big AI leaderboards measure explains why one index never settles it.

The real gains are speed and price. Artificial Analysis measured about 110 output tokens per second against about 57 for Carbon. Each test task took about 21 minutes instead of about 36.

Neon used about 15% more output tokens per task, so the savings come from cheaper tokens, not leaner answers. Its Gemini 4 Neon results page calls the model very verbose.

Who gets it and what it costs

Neon is in the Gemini app and Antigravity for Google AI Pro and Ultra subscribers, and on the paid tier of the Gemini API. The Gemini API pricing page confirms there is no free tier for it.

The API costs $3 per million input tokens and $15 per million output tokens, and the thinking the model does is billed as output. Cached input drops to $0.15. According to Artificial Analysis, that is above Carbon’s launch discount of $2 and $10, which ends on 30 November, and 25% below Carbon’s standard $4 and $20.

Google also says Neon posted its lowest fabrication rate yet in internal tests. We could not check that, and the full safety report for Neon is still listed as coming soon.

What to do this week

If you pay for Gemini Pro or Ultra, take one build job you would hand to Carbon, such as a small game or a messy spreadsheet. Run it on Neon at the default setting. Note whether it finishes, how long it takes and how much of your limit it uses. On the API, try medium or high before xhigh, because every thinking token is billed at the output rate.

Sources

Lars Strother

Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.