Gemini 3.5 Flash: Google's Deprecated Agentic Flash Model, Thinking Levels, Benchmarks and Replacement

What Gemini 3.5 Flash was

Gemini 3.5 Flash is a large language model from Google, released on 19 May 2026 as the next model in the Gemini 3 series. Google DeepMind's model card describes it as a natively multimodal reasoning model built on the Gemini 3 Flash foundation, with thinking levels that trade off quality, cost and latency. Google's developer guide called it its most intelligent Flash model at launch, built for agent and coding work at scale, and released it as a stable, generally available model. It takes text, images, audio, video and PDFs, returns text, and has a 1M-token context window with up to 65,536 output tokens. On LMArena's text leaderboard Gemini 3.5 Flash has 1478 Elo (rank 32, high effort).

Deprecated and routed to Gemini 3.6 Flash

Google has since deprecated Gemini 3.5 Flash. It can no longer be selected in Google AI Studio, and API requests to gemini-3.5-flash are routed automatically to gemini-3.6-flash, which Google describes as a direct replacement with lower token use and lower output pricing. Google asks developers to update their model string to gemini-3.6-flash or to Gemini 3.8 Flash, and points lower-cost, high-volume work to Gemini 3.5 Flash-Lite and Gemini 3.1 Flash-Lite. The Gemini 3.5 Flash guide is kept online for reference. Google DeepMind had named agent workflows, coding and multi-week enterprise processes as its intended uses.

Thinking levels and behaviour

Gemini 3.5 Flash changed the default thinking level from high, used by Gemini 3 Flash Preview, to medium, which Google recommended for most complex code and agent tasks. The low level was significantly improved for shorter code and agent tasks, minimal suits quick chat-like answers, and high allows extended thinking and more tool calls for the hardest problems. The model also keeps its intermediate reasoning across turns of a conversation automatically. Google advises leaving temperature, top_p and top_k at their defaults for all Gemini 3.x models and using thinking_level instead of the older thinking_budget parameter.

Benchmarks Google reported

Google DeepMind's model card lists 76.2% on Terminal-bench 2.1, 55.1% on SWE-Bench Pro, 83.6% on MCP Atlas, 56.5% on Toolathlon and 78.4% on OSWorld-Verified. It also reports 57.9% on Finance Agent v2, an Elo of 1656 on GDPval-AA, 84.2% on CharXiv Reasoning, 83.6% on MMMU-Pro, 40.2% on Humanity's Last Exam and 72.1% on ARC-AGI-2. The same table compares it with Gemini 3 Flash, Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.7 and GPT-5.5, which score higher on several tests, including ARC-AGI-2 and GDPval-AA.

Tools and availability

The API supported code execution, computer use, context caching, file search, function calling, grounding with Google Search and Google Maps, structured outputs and URL context, but not audio generation, image generation or the Live API. It could be used through the Batch API and the Flex and Priority inference options. Google distributed it in the Gemini app, Gemini Enterprise, Google AI Studio, the Gemini API, AI Mode in Google Search and Google Antigravity.

Official Resources