Google DeepMind released Gemini 4 Neon on November 18, the fourth model in its Gemini 4 family in five weeks, and it arrives with a headline result. Neon ranks first on the Artificial Analysis Coding Agent Index with a score of 58.6, ahead of every other model tested, and it does so at a price that undercuts Google’s own flagship once a promotional rate expires.
Neon is built for the everyday work that most developers and knowledge workers actually do, according to the launch post by Koray Kavukcuoglu, Google DeepMind’s chief AI architect. It targets software engineering, analytical reasoning and scientific workflows, and it went live the same day across Google’s API, the Gemini app, its business offering and the Antigravity coding agent.
The model is adapted from Gemini 4 Carbon, the frontier model Google launched in October, and it sits in the middle of the family. Carbon is the deep reasoning tier, Hydrogen is the fast everyday tier, and Neon is the balance between the two.
Gemini 4 Neon by the numbers
| Detail | Gemini 4 Neon |
|---|---|
| Price | $3 per million input tokens, $15 per million output tokens |
| Cached input | $0.15 per million tokens, a 95 percent discount |
| Context window | 1 million tokens, shared between input and output |
| Thinking levels | Four: Low, Medium, High and Extra High, default High |
| Intelligence Index | 61 at xhigh thinking, third overall |
| Coding Agent Index | 58.6, first overall |
| Knowledge cutoff | March 2026 |
| API identifier | gemini-4-neon |
On the Artificial Analysis Intelligence Index, Neon’s 61 puts it third overall, behind only larger and slower frontier models. OpenAI’s GPT-6.2 Astra scores 62, and Claude Fable 5.2 leads at 65. The gap between Neon and the top is small, and Neon is considerably cheaper than both.
Where Neon sits in the Gemini 4 lineup
| Model | Role | Price per million tokens | Intelligence Index |
|---|---|---|---|
| Gemini 4 Carbon | Frontier reasoning | $2 input / $10 output until Nov 30, then $4 / $20 | 60 |
| Gemini 4 Neon | Everyday complex work | $3 input / $15 output, flat | 61 |
| Gemini 4 Hydrogen | Fast and low cost | $1.50 input / $7.50 output | 51 at high, 54 at xhigh |
The pricing table holds a twist. Until November 30, Carbon is actually cheaper per token than Neon, because Carbon is still on its introductory $2 and $10 rate. That promotional rate doubles on December 1 to $4 and $20, which makes Neon 25 percent cheaper than Carbon from that date onward. Google also reports that Neon costs about 17 percent less per task than Carbon at its highest thinking level, and about 32 percent less at the default High setting.
Hydrogen, which launched on November 12 as the cheapest Gemini 4 model, remains the budget option at $1.50 and $7.50. But Neon beats it on every benchmark that matters for coding, and it still costs half as much as Carbon will from December.
The benchmark that matters most
Neon’s headline result is the Coding Agent Index, where its 58.6 puts it first overall. The index measures how well a model works as an autonomous coding agent, doing real repository work rather than answering isolated questions. Neon also scored 65.2 percent on Terminal-Bench and 65.4 percent on CursorBench, beating Carbon and Hydrogen by more than seven points on the latter.
Other results Google cites include second place on the Vals Index at 70.9 percent, second place for steerability on the Agent Arena Index, and 75.8 percent on Agents’ Last Exam, which Google itself describes as level with leading models rather than ahead of them.
A word of caution applies to all of these figures. They come from Google’s own test runs and its chosen benchmark set, and independent verification was incomplete at launch. The Artificial Analysis scores are third-party composite results, which is why they carry more weight than Google’s internal numbers, but broader independent testing will take time.
What changes for Gemini app users
Neon replaces Argon in the Gemini app’s model picker for AI Pro and Ultra subscribers, taking the “Deeper analysis” slot. That is a quiet demotion for Argon, which was one of the original Gemini 4 models, and a sign of how fast Google is iterating. Four models in five weeks is a pace no other AI lab is matching right now.
The model accepts text, images, audio and video as input and replies in text, and it can work across spreadsheets, PDFs, charts, audio and video within a single 1-million-token context window. A normal response can run up to 64,000 tokens, while a background long-decode mode can stretch output up to 1 million tokens, which is useful for very long documents or large code refactors.
There is no free API tier for Neon. Batch and Flex API calls get 50 percent off the standard rate, and existing Gemini 4 integrations can switch to the gemini-4-neon identifier without changing their request format.
What this means for developers
Neon is the model most developers will actually use. Carbon is for the hardest problems, Hydrogen is for the cheapest ones, and Neon is for the repository-scale coding and tool-heavy agent work that fills a normal working day. It is faster than Carbon, roughly twice the throughput by some reports, and it tops the coding agent leaderboard.
The real test will be whether that benchmark lead holds up outside Google’s own harness. If it does, Neon becomes the default choice for coding agents on price and performance, and Google’s strategy of releasing a new Gemini 4 variant every week starts to look less like churn and more like a deliberate squeeze on its rivals.
