Gemini 3 Flash

Overview

Gemini 3 Flash is Google DeepMind's fast, balanced model — released December 17, 2025, it replaced Gemini 2.5 Flash as the recommended mid-tier option. What makes Gemini 3 Flash notable is its benchmark performance: it achieves frontier-level results on PhD-level reasoning tests (GPQA Diamond: 90.4%) and Humanity's Last Exam (33.7% without tools) while maintaining the low latency and cost efficiency expected of a Flash-tier model. This positions it as arguably the best price/performance option in the Gemini lineup for most real-world applications.

If Gemini 3 Pro is Google's answer to the hardest reasoning problems, Gemini 3 Flash is Google's answer for everything else — capable enough for genuine intellectual work, fast enough for production-scale deployment, and priced accordingly.

Specifications

• Developer: Google DeepMind
• Model String: gemini-3-flash (check Google AI docs for versioned string)
• Release Date: December 17, 2025
• Type: Natively Multimodal LLM (text, audio, images, video, code)
• Context Window: 1,048,576 tokens (1M+)
• Access: Google AI Studio, Gemini API, Vertex AI
• Pricing: See Google AI pricing for current rates (positioned below Gemini 3 Pro)

Capabilities

PhD-Level Reasoning at Flash Speed: 90.4% on GPQA Diamond — a benchmark testing graduate-level reasoning across biology, chemistry, and physics — while retaining the latency profile of a Flash-tier model. This is a striking combination.

Humanity's Last Exam: 33.7% without tools on HLE, a benchmark designed to be extremely difficult for AI systems. Scores here compete with larger frontier models.

1M+ Context: Same full context window as Gemini 3 Pro, enabling large-document and codebase analysis without the cost premium.

Native Multimodality: Inherits full Gemini 3 multimodal capability — reasoning across text, images, video, and audio at Flash-tier cost and latency.

Agentic Tasks: Strong performance in tool use and multi-step agentic workflows, making it a practical choice for production AI pipelines where Gemini 3 Pro's cost isn't warranted.

Limitations

For the most demanding reasoning tasks — particularly those requiring the depth of ARC-AGI-2 level novel reasoning — Gemini 3 Pro or 3.1 Pro will outperform it. Flash models inherently trade some depth for speed.

Recent Developments

• December 17, 2025 Launch: Released as the replacement for Gemini 2.5 Flash, with benchmark performance that surprised observers — 90.4% on GPQA Diamond at a mid-tier price point challenges the assumption that frontier reasoning requires top-tier models.
• Best All-Rounder: Multiple independent reviews in early 2026 have noted Gemini 3 Flash as a strong candidate for the best overall price/performance model in the market, competing with Anthropic's Claude Sonnet 4.6.

Last Updated

February 26, 2026