Google launched Gemini 4 Neon, a model for complex programming, analysis, and scientific work that sits between Gemini 4 Carbon and Gemini 4 Hydrogen. Its API costs $3 per million input tokens and $15 per million output tokens; Google says that, at the default reasoning level, Neon delivers performance comparable to Carbon at 32% lower cost per task.
The launch was announced on November 18, and rollout across several Google products began that same day. The company targets it at tasks that require more reasoning than everyday queries but recur often enough for speed and cost per run to matter.
A middle ground between Carbon and Hydrogen
Google describes Carbon as the model for the most demanding, long-running projects, while Hydrogen is designed for fast execution and everyday tasks. Neon aims to fill the middle ground: coding across repositories, debugging, refactoring, business analysis, and multi-step scientific workflows. Its context window is one million tokens, and the model accepts text, images, audio, and video.
Google DeepMind's technical specifications list up to 64,000 synchronous output tokens by default, or up to one million through long-running background decoding. The model is adapted from Gemini 4 Carbon. Google does not present Neon as a replacement for Carbon; instead, it positions it as an alternative for complex, frequent tasks that do not always require the model designed for the longest assignments.
Neon has four reasoning levels: Low, Medium, High, and Extra High. Google sets High as the default; Extra High adds deeper reasoning for difficult engineering, analysis, and science tasks. According to the company, this scale makes it possible to adjust reasoning intensity to the task rather than always using the most expensive setting.
In Google Antigravity, Neon can also act as a planner and delegate routine execution to Hydrogen through Orchestrator. Neon handles the architecture and initial reasoning; Hydrogen receives tasks such as code changes, package updates, and test runs. If it encounters difficulties, the workflow can return the task to Neon so it can review the status and adjust the plan.
Model pricing and availability
The announced API price is $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.15 per million, a 95% discount on the standard input rate. The context window holds one million tokens.
| Item | Gemini 4 Neon |
|---|---|
| Input tokens | $3 per million |
| Cached input tokens | $0.15 per million |
| Output tokens | $15 per million |
| Context window | 1 million tokens |
| Reasoning levels | Low, Medium, High, and Extra High |
In the Gemini app, the model is available to Google AI Pro and Ultra paid subscribers. Google is also rolling it out through the Gemini API in Google AI Studio and Android Studio, in Gemini for businesses, and in Antigravity for Pro and Ultra customers. Vertex AI documentation lists it for verified enterprise customers in global, US, and EU regions. In Antigravity, availability includes the desktop app, CLI, and IDE extensions.
ADSLZone compares Neon's price with those of other Gemini models: it says Neon costs twice as much as Hydrogen and, starting November 30, will be cheaper than Argon and Carbon. Until then, those two models retain promotional prices of $2 per million input tokens and $10 per million output tokens. This temporary offer means that price comparisons depend on the date.

Google's figures compared with Claude and GPT
Google ranks Neon ahead of Claude Opus 5.5 and GPT-6.2 Sol on the Artificial Analysis Intelligence Index: 61 points versus 59 and 57, respectively. The company says this score puts it third in the overall ranking, behind larger and slower models. On the Vals Index, it scores 70.9%, above Claude's 70.1% and GPT-6.2 Sol's 66.1%.
The results do not put it first in every benchmark. On Terminal-Bench 4.1, Google DeepMind's model card records 65.2% for Neon and 65.8% for Claude Opus 5.5; GPT-6.2 Sol scores 61.5%. On FrontierCode 2.0, Neon scores 46.2%, compared with 47.3% for Claude and 43.8% for GPT. On Chartography, which measures the reading of dense charts, Neon reaches 73.5%: below GPT-6.2 Sol's 75.1% and above Claude's 66.3%.
| Benchmark | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 61 | 59 | 57 |
| Vals Index | 70.9% | 70.1% | 66.1% |
| Terminal-Bench 4.1 | 65.2% | 65.8% | 61.5% |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% |
| Chartography | 73.5% | 66.3% | 75.1% |
Google also highlights coding agent benchmarks. Neon scores 58.6 on the Artificial Analysis Coding Agent Index and 56.3% on AutomationBench. On SWE-sweep, it reaches 39% with the Google Antigravity harness, compared with 12.8% on mini-SWE; the model card compares those results with 10.1% and 3.6% for Claude Opus 5.5, and 8.3% and 11.6% for GPT-6.2 Sol, respectively.
The difference between the two SWE-sweep results shows how much the execution environment can influence these measurements. Google attributes part of the higher score to the Antigravity harness. Digital Trends notes that these are comparisons published by Google and that independent evaluations could produce different results.
“To me Neon feels like the right mix of intelligence, speed and usability,”
— Logan Kilpatrick
Working with documents, code, and multimodal data
In the Gemini app, Google recommends Neon for writing, research, and more demanding everyday tasks, such as reviewing receipts or finding changes on a utility bill. It can work with spreadsheets, PDFs, charts, audio, and video in a single request, and draw on Gmail and Docs to add context when needed. Google says it can cross-reference information from multiple sources to improve its analyses and summaries.
For teams working with code, the company highlights editing multiple files, debugging, and refactoring repositories through changes it describes as precise. It also presents the model as suitable for extended tool-based workflows: it can share progress, reduce redundant file reads, and manage context compaction in long sessions.
Google publishes examples of capabilities developed in Antigravity, including a WebGL2 ski simulator built from scratch with 10,200 lines of shaders, 120 Hz physics, and six alpine runs. It also describes a Google spreadsheet with 100,000 cells, more than 500,000 formulas, and six interactive modes, generated from an image without embedded images or Apps Script.
Security, limitations, and access for cyber defense
The model card notes that Gemini 4 Neon may have general limitations, including hallucinations, and sets its knowledge cutoff at March 2026, while warning that it may be less up to date in some areas. Google says it maintains safeguards against serious misuse and prompt injection, and has reduced refusals of legitimate requests compared with earlier models.
In its internal tests, Google reports a 0% cheating rate on Terminal-Bench 4.0, a 2.45% rate of shortcuts in its coding-cheating evaluation, and 0.09% misaligned actions in its agent safety test suite. The model card clarifies that the cited safety evaluations are automated, not human tests or red-team exercises. The full Neon safety report was still listed as forthcoming.
A cyber defense-focused variant will roll out in the coming days to authorized defenders in the Fairwind program. Google says Gemini 4 Neon for cybersecurity scores 76% on the difficult level of AutoBaxBench and 51.26% on WebExploitBench v1.2. It scores 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2; the company describes the latter figure as a 3.4% improvement over Carbon.
The launch makes Neon available to paid subscribers, developers, and enterprise customers through the channels announced by Google. Eligible Fairwind participants will begin receiving access to the cyber defense variant over the coming days.
Sources: ADSLZone, Digital Trends, Google, Google DeepMind, Google Antigravity