Google Gemini 3.2 Flash Leaks: Generates 2200 Lines of Code in One Prompt, Outpacing Claude and GPT

Just before the I/O conference, Gemini 3.2 Flash silently went live and was first spotted by a Reddit user who noticed that the code style generated in Gemini Canvas differed dramatically from the output of the same model in Google AI Studio, suggesting a hidden backend model switch.

The new model appears in the Google Cloud Console as gemini-3.2-flash-lite-live-preview. Community members reported that selecting the “Thinking + Canvas” mode on the Gemini web UI often routes queries to this model.

Gemini 3.2 Flash’s coding power is striking: a single prompt can generate more than 2,200 lines of code, producing interactive SVG UI designs, a full‑featured Three.js project, a PS5‑style blueprint, and even a complete, network‑enabled Windows 98 environment with classic apps, pixel‑perfect taskbar and login experience.

The breakthrough stems from aggressive model distillation and sparsification, which compress the LLM’s core knowledge into a lightweight version without the usual performance collapse associated with smaller models.

Internal benchmarks claim Gemini 3.2 Flash reaches roughly 92 % of GPT‑5.5 performance on core coding and reasoning tasks while cutting inference cost by 15–20× and keeping most query latencies under 200 ms.

Gemini App now integrates third‑party services: Canva for design, Instacart for grocery shopping, and OpenTable for restaurant reservations. Users can issue natural‑language commands such as “design a vintage wedding invitation in rose‑gray and sage‑green” or “add all ingredients from this recipe to my Instacart cart,” and the model executes the tasks directly within the conversation.

The upcoming I/O showcase will also reveal Gemini Spark/Remy (24/7 agents), Gemini Omni (video generation and editing), Gemini 3.5 Flash/Pro (faster, cheaper, lower‑latency), Spark Robin (enhanced visual interaction), and Teamfood (memory‑augmented long‑context). Analysts note that Gemini is positioned to catch up to GPT‑5.5 and surpass Claude Mythos, but it still trails the two major rivals.

I/O 2026 is framed as Google’s “turn‑around” moment, where the company must not only match benchmark numbers but also convince users that it now leads the race toward artificial super‑intelligence.