Gemini 3.2 Flash Revealed: Google’s New Model Beats Its Own Pro in Coding

Shortly before the I/O conference, developers discovered that Google had silently launched Gemini 3.2 Flash on the web, a new model that appears when users select the "Thinking+Canvas" mode in Gemini. A Reddit user first noticed that the code style produced by Gemini Canvas differed dramatically from the same model in Google AI Studio, indicating that the backend had switched to a different model.

The model entry gemini-3.2-flash-lite-live-preview was found in the Google Cloud Console, confirming that the new model was being routed to the web interface. Community members reported that enabling the Thinking+Canvas mode gave a high probability of hitting Gemini 3.2 Flash.

In physical‑simulation 3D tests, Gemini 3.2 Flash demonstrated extraordinary coding power: a single prompt generated over 2,200 lines of Three.js code, intricate SVG graphics, and a fully interactive PS5‑style blueprint. Previously, the Flash model struggled to exceed 400‑500 lines; Gemini 3.2 Flash routinely surpasses 1,000 lines, completing complex tasks in one shot.

Beyond raw line count, the model produced high‑fidelity, interactive assets: a transparent‑balloon effect with realistic lighting, collision feedback, water‑particle effects, and a detailed SVG of a Windows 98 interface complete with a functional browser, classic games, calculator, Paint, Word, and Notepad—all with pixel‑perfect taskbars and login flows.

The breakthrough is attributed to advanced model distillation and sparsification techniques that compress the LLM’s knowledge into a lightweight version without the typical performance drop, reportedly achieving 92% of GPT‑5.5’s code‑generation capability while reducing inference cost by 15‑20× and keeping latency under 200 ms for most queries.

Gemini 3.2 Flash also powers the broader Gemini App, which now integrates third‑party services such as Canva, Instacart, Spotify, WhatsApp, and upcoming integrations with OpenTable and others. Users can ask Gemini to design a vintage‑style wedding invitation in Canva, add ingredients from a recipe directly to an Instacart cart, or reserve a table for eight at a steakhouse—all within a single conversational window.

This positions Gemini as an all‑in‑one AI assistant capable of calling, ordering, designing, and shopping without opening separate apps. The upcoming I/O 2026 event is framed as Google’s chance to prove leadership in the AI race, especially as OpenAI prepares GPT‑5.6 and Anthropic readies its next model. Analysts note that while Gemini 3.2 Flash approaches GPT‑5.5 performance, it still trails Claude Mythos, and the real challenge is convincing users that Google is no longer just chasing competitors but leading the race toward artificial superintelligence.

References: 9to5Google article on Gemini App’s thinking level, a Reddit post exposing the model, and benchmark claims linking Gemini 3.2 Flash to 92% of GPT‑5.5’s code‑generation performance.