Gemini 3 Pro Review: 20%+ Coding Gains and Front‑End SOTA Performance

Performance Metrics

The technical report does not disclose internal architecture, but publishes benchmark results. Compared with Gemini 2.5 Pro, the coding capability of Gemini 3 Pro improves by more than 20 %. Across all evaluated domains the model reaches state‑of‑the‑art (SOTA) scores.

Domain‑specific benchmark tables show the model “crushing” other models, achieving the highest numbers in every category.

Effectiveness

In Google AI Studio the **Build** mode can use Gemini 3 Pro to generate a complete React application quickly.

Examples from the gallery demonstrate:

Product‑website mockups with richer visual styling.

Physics‑collision simulations.

A simple web‑based runner game.

Specific Tests

Prompting the model to write a class‑based chatbot similar to ChatGPT results in a complete, functional implementation in a single pass. The output appears more polished than comparable attempts with other models.

A second prompt asks for a first‑person‑shooter (FPS) game. The initial output resembles an 8‑bit console demo. After refining the prompt, the model produces a more detailed UE‑style scene with functional collision, shooting, and jumping.

New IDE

Google also released an agent‑based IDE named Antigravity. Initial login attempts fail with an authorization error, likely because the service is not yet publicly available or is affected by a Cloudflare outage. Further testing of backend capabilities will be performed after the issue is resolved.

Conclusion

Gemini 3 Pro delivers a noticeable advance in front‑end generation, confirming the expected evolution of the model.

References

[1] https://blog.google/technology/developers/gemini-3-developers

[2] https://aistudio.google.com

[3] https://antigravity.google/product

Code example

[1] https://blog.google/technology/developers/gemini-3-developers
[2] https://aistudio.google.com
[3] https://antigravity.google/product