At Google I/O 2026, Google introduced Gemini 3.5 Flash, a mid-tier multimodal model updated to prioritize speed, improved agentic capabilities, and better visual understanding. Google also noted the new version carries roughly three times the per-token price of its predecessor, Gemini 3 Flash.
Inputs, outputs and technical notes
- Inputs: text, images, audio, and video up to 1 million tokens.
- Outputs: text up to 64,000 tokens; throughput: 204 tokens per second.
- Architecture: a mixture-of-experts transformer.
- Features: adjustable reasoning levels (minimal, low, medium, high), thought preservation (retaining reasoning tokens across multi-turn conversations), and tool use support (computer use not yet available).
According to the model card, Gemini 3.5 Flash is "based on" Gemini 3 Flash, which in turn is based on Gemini 3 Pro. Google disclosed few specifics about parameter counts, training data and methods, and other architectural details. The company says the model was multimodally pretrained on text, code, images, audio and video scraped from the web alongside licensed materials, Google user data, and synthetic data, and later fine-tuned with reinforcement learning on datasets covering multi-step reasoning, problem solving and theorem proving.
Benchmark results
Independent tests report substantial speed and agentic-capability gains over the prior Flash model, while the model generally trails some leading systems on broad intelligence, knowledge and coding metrics.
Notable benchmark outcomes (from Artificial Analysis and other reported measures):
- MMMU-Pro (visual reasoning across academic disciplines): Gemini 3.5 Flash set to high reasoning achieved 84% accuracy, the highest recorded, with Gemini 3.1 Pro Preview at 82%.
- APEX-Agents-AA (long-running agentic tasks): Gemini 3.5 Flash scored 47.1% on first attempt, taking first place and outpacing GPT-5.5 (37.7%).
- GDPval-AA (real-world agentic tasks): Gemini 3.5 Flash at high reasoning reached 1,656 Elo, exceeding Gemini 3.1 Pro Preview (1,314 Elo) but trailing GPT-5.5 set to xhigh (1,769 Elo).
- ARC-AGI-2 (abstract visual reasoning): Gemini 3.5 Flash at high reasoning scored 72.1% on the ARC Prize leaderboard, behind Gemini 3.1 Pro Preview (77.1%) and GPT-5.5 xhigh (85.0%).
- AA-Omniscience (knowledge with penalties for hallucinations): Gemini 3.5 Flash (reasoning setting) scored 23, below Gemini 3.1 Pro Preview (33) and Claude Opus 4.7 at max reasoning (26).
As of May 24, 2026, Arena.ai leaderboards placed Gemini 3.5 Flash ninth in the Text Arena (1,480 Elo) and tenth in the WebDev coding arena (1,506 Elo). Within the Text Arena category breakdowns, the model ranked first in math (1,521 Elo) but 31st in coding (1,507 Elo). Anthropic’s Claude Opus 4.6 and 4.7 occupied the top three positions in both arenas.
Availability and pricing
Gemini 3.5 Flash is available for free via the Gemini app, Google AI Studio, Google Antigravity (within a compute limit that refreshes every five hours up to a weekly limit), and Google Search AI mode. For enterprise and API use, listed prices are $1.50 / $0.15 / $9.00 per million input / cached / output tokens respectively for Gemini Enterprise, Gemini Enterprise Agent Platform and the API.
Google has claimed Gemini 3.5 Flash often runs at less than half the cost of competing models; however, Artificial Analysis found that in its Intelligence Index tests the running cost exceeded that of Gemini 3.1 Pro. The Flash label therefore no longer guarantees a clear cost advantage for agentic workloads.
Other Google I/O AI announcements
Google also announced several related AI updates:
- Antigravity overhaul: Google reworked Antigravity, its AI coding tool, to focus more on managing agents and less on resembling IDEs like Microsoft’s VSCode. The Antigravity command-line version replaces the open-source Gemini CLI.
- Omni family: Google introduced the Omni family of multimodal models, starting with Omni Flash, a lightweight model that can generate video from text, image, audio and video inputs or any combination of them. Omni Flash is available to Google AI Plus, Pro and Ultra subscribers through the Gemini app and Google Flow but is not yet available via API.
- Google Search integration: Gemini 3.5 Flash enables more conversational, chatbot-like queries in Google Search, powers agents to perform online research on users’ behalf, and can shift Search away from its traditional top-ten links toward AI-generated summaries with cited sources.
Why it matters
Gemini 3.5 Flash redefines the Flash tier within Google’s lineup: introduced originally as a smaller, faster tier after Gemini Ultra, Pro and Nano, Flash now functions as a mid-tier multimodal model. Its speed and agentic strengths may justify the higher token costs for developers building multi-turn agents or low-latency applications such as chatbots, search and image/video analysis. At the same time, mixed benchmark results and increased per-token prices mean developers must weigh cost versus performance when choosing to deploy Flash.
