Gemini 3 Flash: Frontier Benchmarks at Lower API Prices

A cheap tier with frontier scores attached

Google released Gemini 3 Flash on December 17, 2025, a month after Gemini 3 Pro, priced at $0.50 per million input tokens and $3 per million output tokens. It shipped as the default model in the Gemini app and AI Mode in Search, and was made available in Google AI Studio, Vertex AI, Gemini Enterprise, Gemini CLI, Android Studio and Google Antigravity. [1]

Google's published scores for Flash include 90.4 percent on GPQA Diamond, 33.7 percent on Humanity's Last Exam without tools, 81.2 percent on MMMU Pro and 78 percent on SWE-bench Verified. Gemini 3 Pro, released the previous month, had scored 91.9 percent on GPQA Diamond and 37.5 percent on Humanity's Last Exam without tools. [1][2]

Atlas interpretation: Those gaps are small enough that a cheap model landing within a few points of a frontier one, on the same benchmarks, is the actual news. Reasoning benchmarks like GPQA Diamond and Humanity's Last Exam are graded on fixed answer keys, so a near miss here is a specific, checkable claim rather than a marketing adjective. [1]

Why the interesting number is the price, not the score

Google's own comparison is not against Gemini 3 Pro but against the prior generation: Flash outperforms Gemini 2.5 Pro on multiple benchmarks while using about 30 percent fewer tokens on average, and Google describes it as roughly three times faster than 2.5 Pro at a lower price. [1]

Atlas interpretation: A year earlier, a Flash-tier release would have been framed against the previous Flash model on cost and speed alone, with a reasoning gap to the Pro tier taken for granted. Framing this one against the previous generation's Pro tier is Google's way of saying the gap between cheap and frontier has narrowed enough that the old distinction, cheap and fast versus capable, no longer describes the product line. That reframes what a Flash-tier release is for: not a discount option beneath the real model, but the version aimed at the high-volume, latency-sensitive and agentic workloads nobody would have routed through a frontier-priced model in the first place. [1]