Google Accidentally Leaked Gemini 3.2 Flash, and It Already Outperforms 3.1 Pro on Coding

Two weeks before Google I/O 2026, an unannounced model called Gemini 3.2 Flash showed up in the wild. No blog post, no keynote, no press release. Users just found it, and what they found is seriously impressive.

Gemini 3.2 Flash: a double leak

On May 5, a Reddit user noticed the iOS Gemini app's model selector cycling through versions over 24 hours, eventually landing on 3.2 Flash. Around the same time, the model was spotted running silent benchmarks on Arena AI (formerly known as LM Arena). Google has not confirmed anything.

The iOS build also revealed a fresh interface called "Liquid Glass" with a pill-shaped prompt box and an animated gradient background. The dedicated Thinking mode is gone, replaced by a global reasoning toggle across all model tiers.

Gemini 3.2 Flash outperforms 3.1 Pro on coding

Early Arena tests tell a compelling story. The model excels at SVG generation, interactive 3D coding, and animation processing. A viral test asked it to produce an animated ASCII city skyline: 3.2 Flash delivered in under two minutes while 3.1 Pro produced broken code.

Leaked API Studio metadata suggests pricing at $0.25 per million input tokens and $2.00 per million output tokens. That output price is 33% cheaper than Gemini 3 Flash. A Flash-tier model beating a Pro on creative coding while costing less is the real headline here.

Google I/O timing and competitive pressure

Google I/O runs May 19–20 at Shoreline Amphitheatre. Polymarket traders currently give an 83% probability that Gemini 3.2 launches before May 31. The 3.2 versioning instead of 3.5 signals a shift toward faster, incremental releases.

With OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 both pushing hard on coding capabilities, Google needs every edge. Whether this leak was an accident or a calculated tease, the takeaway is the same: the price-to-performance ratio on Flash models just got absurdly competitive. Developers building on the Gemini API should pay attention.