Google releases Gemini 4 Neon for developers

The new superintelligence (SI) model is available through an API and appeared in Google Cloud’s Vertex AI Model Garden, giving developers another Gemini option to assess.
In this article
Google released Gemini 4 Neon on 18 November 2026, with access through an application programming interface (API), which lets developers connect software to the model. Neon also appeared in Google Cloud’s Vertex AI Model Garden.
ArtificialWatch reported that leak accounts on X claimed Neon had reached Google AI Studio, while some users said access was limited to certain accounts.
The release follows Gemini 4 Carbon, which Google announced on 28 October 2026, with initial access limited to a subset of cyber defenders through the Fairwind Program.
Google says Neon completes complex software engineering, knowledge-work and science tasks quickly. It says the model reasons more deeply than Gemini 4 Hydrogen and offers quality close to Carbon on coding, reasoning and agentic tasks, in which the model acts with a degree of autonomy. ArtificialWatch said those comparisons were Google’s description and that no independent benchmark scores for Neon were available at launch.
Developer prices
API prices are based on tokens, the units used to process input and output. Carbon and Argon’s introductory prices run through 30 November; their higher prices take effect on 1 December.
| Model | Input | Output |
|---|---|---|
| Gemini 4 Neon | $3 | $15 |
| Gemini 4 Hydrogen | $1.50 | $7.50 |
| Gemini 4 Carbon and Argon | $2.00 introductory / $4.00 afterward | $10.00 introductory / $20.00 afterward |
Inputs and limits
Neon accepts text, images, video, audio, PDFs and Office and OpenDocument files. It supports Computer Use, code execution and file search, as well as grounding through Google Search and Maps. Grounding connects responses to information from those services. Neon cannot generate images or audio.
ArtificialWatch said Neon has a 1M-token context window shared by input, thinking and output. A context window is the amount of information the model can process together. The outlet said it can produce up to 64K output tokens per segment, or up to 1M using Long Decode.
The Decoder expected benchmarks and further details within a few hours of its November 18 report.
Follow AveNews
Earlier in Technology
SoftwareUS cyber agency adds four exploited software flaws to catalog
The Cybersecurity and Infrastructure Security Agency says attackers are exploiting vulnerabilities in Synology DiskStation Manager, BerriAI LiteLLM, Apache Axis and Vanderbilt University’s REDCap, prompting remediation deadlines.
3 min
SuperintelligenceMicrosoft opens Web IQ search service to Azure customers
The service gives developers a way to supply superintelligence (SI) agents with web search results. Microsoft opened it more broadly at Ignite on 17 November 2026, after introducing it at Build in June 2026.
2 min
SoftwarePlayStation Network disruption affects gaming and social services
Sony acknowledged problems launching games, apps and network features on 18 November. Its status page later showed no incidents in France or the rest of Western Europe, while outage messages remained for Russia.
2 min
