Gemini 3.5 Flash-Lite

What it is

Gemini 3.5 Flash-Lite is the fastest model in Google's Gemini 3.5 series. It was released on 21 July 2026 and is aimed at low-latency, high-throughput work — jobs where many requests arrive at once and each one has to come back quickly, such as agentic search and document processing^[1][2]. Agentic search means a model that issues searches, reads the results, and searches again on its own; document processing means reading, extracting from, and summarizing large numbers of files. Both involve many round trips, which is why speed matters as much as raw capability.

In Google's naming, the Lite suffix marks the smaller, cheaper member of a model family, trading some capability for speed and a lower price.

Speed and price

The benchmarking firm Artificial Analysis measured the model at 350 output tokens per second, faster than Gemini 3.5 Flash^[1]. Output tokens per second describes how quickly a model generates text once it starts replying; higher figures matter where a system makes many short calls in sequence.

Standard pricing is $0.30 per 1 million input tokens and $2.50 per 1 million output tokens^[1][3]. Input tokens are the text a caller sends in — instructions, documents, search results — and output tokens are what the model generates; splitting the two prices reflects the different cost of reading a large prompt and writing a long answer. Google also listed discounted batch and Flex rates of $0.15 per 1 million input tokens and $1.25 per 1 million output tokens, and priority pricing of $0.54 per 1 million input tokens for text, image and video and $4.50 per 1 million output tokens^[3].

Reported scores

Google's reported figures covered several kinds of evaluation. The model scored 54% on Terminal-Bench 2.1, against 31% for the comparison Google cited, and 72.2% on GDM-MRCR v2, a test of handling long contexts — retrieving information from very long inputs. On GDPval-AA v2 it scored 1140; Google said the model significantly outperformed 3.1 Flash-Lite^[1].

Against Gemini 3 Flash, the model scored 54.2% versus 49.6% on SWE-Bench Pro, which tests software-engineering work, and 74.0% versus 65.1% on OSWorld-Verified, which tests an agent's ability to carry out tasks through a computer interface^[1].

What developers get

Developers can choose a thinking level from minimal to high^[1]. Thinking levels control how much internal reasoning the model does before answering: a low setting returns answers faster and cheaply for simple lookups, while a high setting spends more time working through harder problems. Computer use is available as a built-in tool, so an application can let the model drive a graphical interface directly rather than only exchanging text^[1].

The model is available in the Gemini API, Antigravity, Gemini Enterprise and the Gemini app, and is rolling out in Google Search^[1].

Where it sits in the lineup

Google designated the model as the replacement for gemini-3.1-flash-lite, and as the auto-routing target for gemini-2.5-flash from 1 December 2026^[2]. Both are moves inside the API: when an older model name is retired, code that still calls it has to point somewhere else, and auto-routing means requests addressed to gemini-2.5-flash are directed to Flash-Lite instead. For developers running high-volume, cost-sensitive systems, that makes the Lite model the default path forward rather than one option among several.