Gemini 3.5 Flash
A multimodal reasoning model from Google DeepMind with adjustable thinking levels
Gemini 3.5 Flash is a natively multimodal reasoning model from Google DeepMind, published as a model card on 19 May 2026. It builds on the Gemini 3 Flash foundation and adds thinking levels that let users trade off quality against cost and latency, with a context window of up to one million tokens.
- Published 19 May 2026 as part of the Gemini 3 series
- Built on the Gemini 3 Flash reasoning foundation
- Thinking levels let users balance quality, cost and latency
- Handles text, images, audio and video with up to 1M tokens of context
- Scored 76.2% on Terminal-bench 2.1 agentic coding as of May 2026
What it is
Gemini 3.5 Flash is a natively multimodal reasoning model from Google DeepMind, described as the next step in the Gemini 3 series. Its model card was published on 19 May 2026. It accepts text, images, audio and video, with a token context window of up to one million. Google DeepMind positions it for users, developers and enterprises, naming agentic workflows, coding tasks and multi-week enterprise processes as target uses. It sits on the Gemini 3 Flash reasoning foundation and adds thinking levels, which control the balance between answer quality, cost and response time. Parameter counts and training details are not given on its own card.
How it came about
Google DeepMind published the Gemini 3.5 Flash model card on 19 May 2026. The model is presented as an iteration within the Gemini 3 series rather than a fresh design, built directly on Gemini 3 Flash. Its card points readers to the Gemini 3 Flash card for architecture, training data, data processing, hardware and software, and to the Gemini 3.1 Pro card for frontier safety assessment and risk mitigation. The company says it continues to refine its internal evaluations, so results in this card are not directly comparable with earlier Gemini cards.
Inside the model
Google DeepMind gives no separate architecture or training description for Gemini 3.5 Flash. The card states it is based on the Gemini 3 Flash reasoning foundation and refers readers to that model's card for architecture, training dataset, data processing, hardware and software. What is new is a set of thinking levels that let users tune the mix of quality, cost and latency. Inputs are text strings, images, audio and video files, handled natively in one model, with a context window of up to one million tokens.
What it can do
Google DeepMind reports results as of May 2026 across reasoning, coding, agentic tool use, multimodal tasks, multilingual performance and long context. On Terminal-bench 2.1 agentic coding it scored 76.2%, and 55.1% on SWE-Bench Pro (Public). It reached 83.6% on the MCP Atlas multi-step workflow test and 78.4% on OSWorld-Verified computer use. CharXiv Reasoning came in at 84.2% and MMMU-Pro at 83.6%. On Humanity's Last Exam it scored 40.2% and on ARC-AGI-2 72.1%. Long-context MRCR v2 (8-needle) was 77.3% at 128k and 26.6% at 1M pointwise.
How to use it
The model is distributed through the Gemini Enterprise Agent Platform and made available to downstream providers by API, under relevant terms of use. Google DeepMind says no special hardware or software is needed to use it. Terms differ by channel: the Gemini API Additional Terms of Service cover AI Studio and the Gemini API, while the Google Cloud Platform Terms of Service cover the Gemini Enterprise Agent Platform. Model API instructions and a quickstart are published alongside.
Safety and testing
Google DeepMind reports automated internal safety evaluations, not human ratings. Against Gemini 3 Flash, text-to-text safety fell 3.9% and multilingual safety 2.6%, image-to-text safety was unchanged, tone improved 8.9%, and unjustified refusals rose 0.8% on non-egregious cases. Manual review found flagged losses were mostly false positives or not egregious. Specialist red teams outside the model team tested it, and child safety launch thresholds were met. Frontier safety work ran on Gemini 3.1 Pro, which reached no Critical Capability Levels; the company says 3.5 Flash adds no meaningful new capability there and stays below the cyber level.
Questions people ask about Gemini 3.5 Flash
What is Gemini 3.5 Flash?
It is a natively multimodal reasoning model from Google DeepMind, part of the Gemini 3 series. Published on 19 May 2026, it builds on the Gemini 3 Flash foundation and adds thinking levels that control the trade-off between quality, cost and latency.
What can it take as input?
Text strings such as questions, prompts or documents to summarise, plus images, audio and video files. The token context window runs to one million tokens, so long documents and long video or audio inputs can be handled in one go.
How do I get access to it?
Google DeepMind offers it through the Gemini Enterprise Agent Platform and to downstream providers by API. No special hardware or software is required. AI Studio and the Gemini API fall under the Gemini API Additional Terms of Service, while the platform falls under Google Cloud Platform terms.
How does it compare with Gemini 3 Flash?
On the May 2026 figures it scored higher than Gemini 3 Flash on every listed benchmark, for example 76.2% versus 58.0% on Terminal-bench 2.1 and 84.2% versus 80.3% on CharXiv Reasoning. Safety results were mixed: tone improved while text-to-text and multilingual safety scores fell slightly.
Does it raise new frontier safety concerns?
Google DeepMind says no. Frontier safety testing was run on Gemini 3.1 Pro, which reached no Critical Capability Levels in its Frontier Safety Framework. The company judges that 3.5 Flash adds no meaningful new capability there, and extra cyber testing kept it below the cyber level.
Good reading and sources
There is no English Wikipedia article on Gemini 3.5 Flash yet, so the facts on this page come from the sources listed above: deepmind.google/models/model-cards/gemini-3-5-flash/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.