Google's Logan Kilpatrick on Gemini 3 Flash: faster, cheaper, and better than Gemini 3 Pro in key benchmarks
Dec 17, 2025 with Logan Kilpatrick
Key Points
- Google's Gemini 3 Flash outperforms Gemini 3 Pro on key benchmarks like ARC-AGI 2 because the smaller model received a month of post-training improvements Pro lacked at launch, not due to architectural superiority.
- Gemini 3 Flash costs less and runs faster than Gemini 2.5 Pro while matching or exceeding its quality, driving higher retention and engagement when Google switched its own AI Studio product to the new model.
- Google plans to release Gemini 3 Flash Lite in early 2025 and is rebasing specialized models for image generation, text-to-speech, and computer use onto the new generation to inherit performance gains automatically.
Summary
Gemini 2.5 Flash's successor arrives with a counterintuitive result: Gemini 3 Flash outperforms Gemini 3 Pro on several benchmarks, including ARC-AGI 2. The explanation is straightforward rather than architectural. Because Flash is a smaller model, Google's research team was able to apply roughly a month's worth of post-training innovations that hadn't yet been incorporated into Pro at the time Pro shipped. The gap reflects timing, not a fundamental capability inversion.
Logan Kilpatrick, who leads developer relations for Google's AI products, frames the Flash launch primarily as a cost and speed story for the developer ecosystem. Customers currently on Gemini 2.5 Pro, paying between $1.25 and $2.50 per million input tokens and $10 to $12 per million output tokens, can migrate to 3 Flash and receive a product that is faster, cheaper to run, and comparable or superior in quality across most tasks. Kilpatrick describes an internal test on Google's own AI Studio vibe-coding product where switching to 3 Flash drove higher retention, more builds, and greater engagement with no engineering effort beyond changing the model in a dropdown.
The broader model family roadmap points toward a Gemini 3 Flash Lite arriving in early 2025, continuing the tiered structure of Pro, Flash, Flash Lite, and an on-device Nano variant. Kilpatrick confirmed these smaller models are in progress but noted that trickle-down innovation takes time.
Kilpatrick also flagged that 3 Flash and 3 Pro now serve as the foundation for Google's specialized model variants, including image generation, text-to-speech, live audio, and computer use models. As those checkpoints rebase to the new generation, they inherit performance improvements automatically.
On Google Workspace's agent builder, Kilpatrick offered a candid internal-user perspective. As a Google employee restricted from using third-party AI tools in a work context, he relies on the Workspace agent builder daily for email triage and workflow automation. He called it genuinely useful and positioned it as the first chapter of a broader AI-native Workspace story, drawing a parallel to how Google Search's AI mode, led by Robbie Stein, evolved from a speculative bet into a competitive frontier product at search scale.
Google goes offline starting the following Monday, with Kilpatrick flagging a few more releases before the break and describing January's roadmap as, in his words, "ridiculous."