Gemini 3 Flash Preview is a high speed, high value thinking model designed for agentic workflows, multi turn chat, and coding assistance. It delivers near Pro level reasoning and tool use performance with substantially lower latency than larger Gemini variants, making it well suited for interactive development, long running agent loops, and collaborative coding tasks. Compared to Gemini 2.5 Flash, it provides broad quality improvements across reasoning, multimodal understanding, and reliability.
The model supports a 1M token context window and multimodal inputs including text, images, audio, video, and PDFs, with text output. It includes configurable reasoning via thinking levels (minimal, low, medium, high), structured output, tool use, and automatic context caching. Gemini 3 Flash Preview is optimized for users who want strong reasoning and agentic behavior without the cost or latency of full scale frontier models.
Evaluations
11
across 10 benchmarks
Latency
39.7s
Context length
1M
tokens
Cost
$0.50 · $3
input · output per 1M tokens
Key takeaways
The Gemini 3 Flash Preview (high) model is a Transformer-based model designed for agentic workflows, multi-turn chat, and coding assistance, optimized for strong reasoning and agentic behavior with lower latency and cost than larger models. It supports multimodal inputs (text, images, audio, video, PDFs) with text output, configurable reasoning via thinking levels, structured output, tool use, and automatic context caching.
The model demonstrates strong logical reasoning and step-by-step problem-solving, achieving 93.33% accuracy on the AIME 2024 problems involving complex algebra, number theory, geometry, and combinatorics. Its performance was nearly perfect, with only one error in a highly abstract functional equation. However, it shows sensitivity to strict output formatting.
A significant limitation is its performance in function calling, with only 53.55% accuracy on the Berkeley-Function-Calling-v3 benchmark. It struggles with parameter extraction, incorrect tool selection, multi-step reasoning, and discerning when no tool is applicable, often missing required parameters or misinterpreting intent.
On the Humanity's Last Exam (HLE) benchmark, the model achieved a low accuracy of 28.18%, struggling with nuanced interpretations of medical scenarios, precise numerical calculations, and frequently misunderstanding the core question in multi-modal tasks. It also had difficulty with fine-grained pattern recognition in abstract domains like chess.
The model performs inconsistently across subjects, sometimes excelling in highly specialized areas while failing in seemingly simpler ones on the HLE. Its performance suggests a critical need for improvement in contextual reasoning, fine-grained detail extraction, and robust numerical precision for high-stakes academic benchmarks.
