Gemini 3.5 Flash

What it is

Gemini 3.5 Flash is a large language model released by Google and the first member of the Gemini 3.5 family. Google announced it at Google I/O on 19 May 2026, in a post by Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals and Noam Shazeer^[1]. The company described the model as built for agents and coding, and said it showed frontier performance on long-horizon tasks, meaning work that unfolds over many steps rather than a single question and answer^[1].

What Google said it could do

Google said Gemini 3.5 Flash outperformed Gemini 3.1 Pro, an earlier model in the same line, on the coding and agentic benchmarks it listed^[1]. The reported scores were 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, 83.6% on MCP Atlas and 84.2% on CharXiv Reasoning^[1]. Benchmarks of this kind generally test how well a model uses tools and a terminal, carries out multi-step tasks, and reads charts and figures; a score describes behaviour on that test set and not on ordinary use.

Google also said the model's output speed was about four times that of other frontier models, measured in tokens per second^[1]. Output speed matters for agentic work because an agent may need many model calls to finish one task, so a faster model shortens the whole run.

Where it shipped

The model launched generally available on the day it was announced. It became the default model in the Gemini app and in AI Mode in Search for all users globally, and it was also offered in Google Antigravity, the Gemini API, Android Studio, Gemini Enterprise Agent Platform and Gemini Enterprise^[1][2]. It powered Gemini Spark, a 24/7 personal agent that Google introduced the same day^[1]. Making a model the default in consumer products is a distribution decision as much as a technical one: it puts the model in front of people who never open a developer console or choose a model by name.

The rest of the family

Gemini 3.5 Flash arrived before the rest of its family. At announcement, Gemini 3.5 Pro was already in internal use at Google, and the company planned to roll it out the following month^[1]. In July 2026 Google said 3.5 Pro was still being tested with partners^[4].

A cybersecurity variant, Gemini 3.5 Flash Cyber, was fine-tuned for finding and fixing vulnerabilities and was offered to governments and trusted partners through a CodeMender pilot^[4]. It is the same base model pointed at a narrower job: reading code and looking for security flaws rather than answering general requests.

Where it sits in the line

The model was superseded within the same year. Gemini 3.6 Flash, released on 21 July 2026, used 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index^[4]. gemini-3.5-flash was removed from AI Studio, and API requests naming it are automatically routed to gemini-3.6-flash^[3][4]. Google's Gemini API deprecations page said the gemini-3.5-flash endpoint is not shutting down and that its API requests are automatically routed to gemini-3.6-flash^[3]. In practice, software that still names the older model keeps running, but the answers come from the newer one.

Taken together, the release shows the shape of Google's model releases at that point: one fast model first, made the default across consumer surfaces, with a larger sibling and a specialised security variant following, and a replacement arriving a couple of months later while the old endpoint stays alive in name only.