Google released Gemini 4 Neon on 18 November, a model adapted from Gemini 4 Carbon for coding, analysis and other multi-step work. At its default setting it scores 60 on the Artificial Analysis Intelligence Index, the same as Carbon, at what Google says is a 32% lower cost per task. Unlike Carbon, AI Pro subscribers get it in the Gemini app, and developers pay $3 per million input tokens and $15 per million output tokens.
AI Pro subscribers can pick Neon in the Gemini app, while Carbon stays with Ultra
Google AI Pro and Ultra subscribers can choose Neon from the model menu in the Gemini app. The app’s Tasks feature can now run on it as well. Carbon, which Google positions for the longest and hardest projects, still needs an Ultra subscription in the app, VentureBeat reports, citing Google’s Logan Kilpatrick.
Neon is on the paid tier of the Gemini API, in Google AI Studio and in Android Studio. AI Pro and Ultra subscribers can also try Gemini 4 Neon in Google AI Studio. In Antigravity, Google’s coding environment for AI Pro and Ultra subscribers, Neon can direct several Hydrogen agents that work in parallel. Companies on Gemini for business can select it in the model menu or call it through the API.
The model card and Google’s post give these specifications for developers:
- Model ID: gemini-4-neon
- Context window: 1 million tokens, with up to 64,000 output tokens per normal request
- Price: $3 per million input tokens, $0.15 for cached input and $15 for output
- Knowledge cutoff: March 2026
Google’s third place uses the xhigh setting, which more than doubles the output
Neon has four thinking levels: low, medium, high and xhigh. High is the default. Google’s two headline claims come from different levels. The 32% saving against Carbon is measured at high. The third place on the Artificial Analysis index, with a score of 61, is measured at xhigh.
That extra point is expensive. Artificial Analysis measured just under 15 minutes per index task at high and 21.5 minutes at xhigh. Across the whole index, xhigh wrote 283 million output tokens against 123 million at high, and output tokens are what you pay most for. Google says Neon at xhigh still costs at least 36% less per task than rival models at their highest settings. It does not name those models.
From 1 December, Neon costs less per token than Carbon
Carbon’s introductory API price ends on 30 November. From 1 December, Carbon costs $4 per million input tokens and $20 per million output tokens, so Neon becomes the cheaper of the two per token. That comes on top of Google’s claim that Neon needs fewer tokens per task. The $0.15 price for cached input is a 95% discount, which matters for agents that send the same instructions and documents again and again.
On the Artificial Analysis leaderboard of 19 November, only Claude Fable 5.2 and GPT-6.2 Astra score higher than Neon at xhigh. The scores of Neon, Carbon and Claude Opus 5.5 sit within two points, so price and speed decide more than the ranking does:
| Measured | Gemini 4 Neon | Gemini 4 Carbon | Claude Opus 5.5 |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 60 (high), 61 (xhigh) | 60 (high) | 59 (max) |
| Input price per million tokens | $3 | $2, then $4 from 1 December | $4 |
| Output price per million tokens | $15 | $10, then $20 from 1 December | $20 |
The scores are from Artificial Analysis, an independent benchmarking firm. The Gemini prices are Google’s, and the Opus price is Anthropic’s list price as reported by VentureBeat. Most other figures in Google’s post, such as 56.3% on Zapier’s AutomationBench, were reported by Google itself.
Google has not yet said where Neon processes data or published its full safety report
Google names no countries for the launch. The Netherlands is on Google’s list of countries where the Gemini API works. The post mentions no exception for Neon, so Dutch developers can call it today. Google does not name a data region for Neon, and the post does not mention Gemini Enterprise Agent Platform, its developer platform on Google Cloud.
The model card says the full safety and security report for Neon will follow in the coming days. If your AI policy asks for a vendor’s safety documentation before a model goes into production, that document is still missing.
For teams already on Google, Neon makes Carbon the exception
For organisations that already pay for Google AI Pro, Gemini for business or the Gemini API, Neon is now the Gemini model to test first. It scores like Carbon at its default setting, Google says it costs less per task, and from December it costs less per token as well. Carbon remains the choice for the longest projects, but most agent and coding work will not need it. Testing inside a contract you already have is also quicker than adding a vendor, with a new data processing agreement and new access rights to arrange.
The number to compare is the cost of a finished task, not the price per token. Take one real task your team now runs on Carbon, Claude Opus 5.5 or GPT-6.2 and run it on Neon at high. Record the cost, the time and how often a person has to fix the result. Move to xhigh only where high falls short. Keep client and personal data out of the test until Google says where Neon processes it.
For more information, visit the official announcement of Gemini 4 Neon on the Google blog.