جیمنی 2.5 فلیش لائٹ

Google's most cost-efficient multimodal model of its generation, offering the fastest performance for high-frequency, lightweight tasks. It is aimed at high-volume classification, simple data extraction and very low-latency applications where budget and speed are the primary constraints. Thinking is off by default, which is what makes it cheap.

Google described this model as pushing the frontier of intelligence per dollar, and everything about it follows from that sentence. It was the cheapest and fastest model in the 2.5 family at general availability, and the post pairs the price with a 40% cut to audio input pricing that had applied during the preview - a detail that matters if the workload is transcript or call data rather than text.

The speed claim is made against specific predecessors: Google reports lower latency than both the previous generation's Flash-Lite and its full Flash model across a broad sample of prompts, and higher all-round quality than the previous Flash-Lite on coding, mathematics, science, reasoning and multimodal understanding. The lab is careful to say this is a balance for latency-sensitive work such as translation and classification, not a claim about hard problems.

Being the cheap tier does not mean being a stripped one. It carries the same million-token context window as the rest of the family, controllable thinking budgets, and the native tools - grounding with Google Search, code execution and URL context - so a high-volume pipeline does not have to fall back to a larger model just to fetch a page or run a snippet.

The deployments Google chose to name are a good description of the shape of work it fits: summarising satellite telemetry in orbit, where the lab reports a 45% latency reduction and 30% lower power draw against the operator's previous model; planning and translating video content across more than 180 languages; and processing long product videos into documentation by extracting thousands of frames. All of them are the same job done a very large number of times.

What it is not for is anything demanding. Reasoning is off by default here and has to be turned on deliberately, and Google's own framing puts Flash above it for everyday tasks and Pro above that for coding and complex work. Choosing it is choosing throughput and cost over headroom.