Gemini 3 Pro Preview arrives on Vercel’s AI Gateway and v0
Google’s Gemini 3 Pro Preview model is now live on the Vercel AI Gateway, the AI SDK, and v0. Vercel says it has been stress-testing the model across its stack for several weeks, with a particular focus on coding, multimodal reasoning, and tool use.
In Vercel’s internal Next.js evals, Gemini 3 Pro Preview registered an almost 17% improvement in correctness over its predecessor. That performance places it in the top two models on the leaderboard, where it is tied with openai/gpt-5-codex at a 42% success rate. It also outperforms anthropic/claude-opus-4.1, zai/glm-4.6, and moonshotai/kimi-k2-turbo in consistency and reasoning quality.
Model | Total Evals | Success Rate | Avg Duration | Total Tokens |
| 50 | 42% | 42.80s | 186,082 |
| 50 | 42% | 44.85s | 309,263 |
| 50 | 40% | 20.36s | 106,177 |
| 50 | 40% | 29.47s | 165,810 |
| 50 | 38% | 4.13s | 82,567 |
Where the model shines
Vercel’s testing highlights Gemini 3 Pro Preview’s strength in frontend generation and multi-turn agent workflows. The model follows developer-level prompts with precision, producing stable, well-structured UI components. Its 1M context window allows it to maintain fidelity across long generation chains, and tests within v0 show gains in both generation speed and output quality.
In Vercel’s Next.js evals, Gemini 3 Pro Preview tied with openai/gpt-5-codex at a 42% success rate, surpassing anthropic/claude-opus-4.1, zai/glm-4.6, moonshotai/kimi-k2-turbo, and others in consistency and reasoning quality.
Testing it immediately
Developers can try the model without any setup via the AI Gateway Playground, which offers a simple chat interface for instant testing. The model string can then be copied directly into the AI SDK. Alternatively, the AI SDK playground allows side-by-side comparisons with other models.
Gemini 3 Pro Preview is available today on AI Gateway with support for both Google AI Studio and Google Vertex AI. v0 users can also select it from the model picker; Vercel notes it is faster and more detail-oriented than the original v0 models, making it well suited for frontend and multi-step AI agents.
import { streamText } from 'ai'
const result = streamText({
model: "google/gemini-3-pro-preview",
prompt: "Create a neural fluid engine simulation in a self-contained HTML"
});
Building custom integration
For teams that want their own interface, the AI SDK and AI Gateway can be combined with any framework to build custom solutions on the same deployment-aware infrastructure and security stack that powers production apps on Vercel. This release is part of the broader Vercel AI Cloud offering, which targets developers building with the latest frontier models.
