Gemini 3.5 Flash, published in May 2026, is the next iteration within the Gemini 3 series, characterized as a highly-capable, natively multimodal, reasoning model. It builds upon the Gemini 3 Flash reasoning foundation, incorporating "thinking levels" to dynamically manage the trade-off between quality, cost, and latency. The model accepts diverse inputs, including text strings (e.g., questions, prompts, documents for summarization), images, audio, and video files, supporting a substantial context window of up to 1 million tokens. Its primary output modality is text, with a maximum output length of 64,000 tokens. Architectural details, training datasets, training data processing, hardware specifications, and software implementations are largely inherited from or referenced to the Gemini 3 Flash model card.

Gemini 3.5 Flash is distributed across multiple channels, including the Gemini App, Gemini Enterprise App, Gemini Enterprise Agent Platform, Google AI Studio, Gemini API, Google Search AI Mode, and Google Antigravity. It is available to downstream providers via an API, requiring no specific hardware or software beyond internet access and adherence to relevant terms of service.

The model underwent comprehensive evaluation across a spectrum of benchmarks as of May 2026, encompassing reasoning, coding, agentic tool use, multimodal capabilities, multilingual performance, and long-context understanding. Key evaluation results demonstrating its performance, particularly relative to prior Gemini models and competitors, are as follows:

• Coding:
  • Terminus-2 harness (Agentic terminal coding): 76.2% (vs. 58.0% for Gemini 3 Flash, 70.3% for Gemini 3.1 Pro, 66.1% for Claude Sonnet 4.6, 78.2% for GPT-5.5).
  • SWE-Bench Pro (Public) (Diverse agentic coding tasks, single attempt): 55.1% (vs. 49.6% for Gemini 3 Flash, 54.2% for Gemini 3.1 Pro, 64.3% for Claude Sonnet 4.6, 58.6% for GPT-5.5).
• Agentic MCP (Multi-step workflows using MCP): 83.6% (vs. 62.0% for Gemini 3 Flash, 78.2% for Gemini 3.1 Pro, 69.5% for Claude Sonnet 4.6, 79.1% for Claude Opus 4.7, 75.3% for GPT-5.5).
• Toolathlon (Real-world general tool use): 56.5% (vs. 49.4% for Gemini 3 Flash, 55.6% for GPT-5.5).
• UI Control OSWorld-Verified (Agentic computer use): 78.4% (vs. 65.1% for Gemini 3 Flash, 76.2% for Gemini 3.1 Pro, 72.5% for Claude Sonnet 4.6, 78.0% for Claude Opus 4.7, 78.7% for GPT-5.5).
• Expert tasks:
  • Finance Agent v2 (Financial analysis and decision-making): 57.9% (vs. 42.6% for Gemini 3 Flash, 43.0% for Gemini 3.1 Pro, 51.0% for Claude Sonnet 4.6, 51.5% for Claude Opus 4.7, 51.8% for GPT-5.5).
  • GDPval-AA (Economically valuable knowledge work): Elo 1656 (vs. 1204 for Gemini 3 Flash, 1314 for Gemini 3.1 Pro, 1676 for Claude Sonnet 4.6, 1753 for Claude Opus 4.7, 1769 for GPT-5.5).
• Multimodal:
  • CharXiv Reasoning (Information synthesis from complex charts, no tools): 84.2% (vs. 80.3% for Gemini 3 Flash, 83.3% for Gemini 3.1 Pro, 72.4% for Claude Sonnet 4.6, 82.1% for Claude Opus 4.7, 84.1% for GPT-5.5).
  • MMMU-Pro (Multimodal understanding and reasoning, no tools): 83.6% (vs. 81.2% for Gemini 3 Flash, 80.5% for Gemini 3.1 Pro, 74.5% for Claude Sonnet 4.6, 75.2% for Claude Opus 4.7, 81.2% for GPT-5.5).
  • Blueprint-Bench 2 (Agentic spatial reasoning, Normalized score): 33.6% (vs. 0.0% for Gemini 3 Flash, 26.5% for Gemini 3.1 Pro, 6.7% for Claude Sonnet 4.6, 24.5% for Claude Opus 4.7, 36.2% for GPT-5.5).
• Long context:
  • MRCR v2 (8-needle) (Long context performance): 128k (average) 77.3% (vs. 67.2% for Gemini 3 Flash, 84.9% for Gemini 3.1 Pro, 84.9% for Claude Sonnet 4.6, 59.3% for Claude Opus 4.7, 94.8% for GPT-5.5). 1M (pointwise) 26.6% (vs. 22.1% for Gemini 3 Flash, 26.3% for Gemini 3.1 Pro).
• Reasoning:
  • Humanity’s Last Exam (Academic reasoning, full set, text + MM): 40.2% (vs. 33.7% for Gemini 3 Flash, 44.4% for Gemini 3.1 Pro, 33.2% for Claude Sonnet 4.6, 46.9% for Claude Opus 4.7, 41.4% for GPT-5.5).
  • ARC-AGI-2 (Abstract reasoning puzzles): 72.1% (vs. 33.6% for Gemini 3 Flash, 77.1% for Gemini 3.1 Pro, 58.3% for Claude Sonnet 4.6, 75.8% for Claude Opus 4.7, 84.6% for GPT-5.5).

Gemini 3.5 Flash is intended for use by developers, users, and enterprises, particularly excelling in agentic workflows, coding tasks, and multi-week enterprise processes. Known limitations and acceptable usage guidelines refer to the Gemini 3 Flash model card.

Regarding ethics and content safety, the evaluation approach and safety policies are consistent with those detailed in the Gemini 3 Flash model card. Internal automated safety evaluations during development showed the following performance relative to Gemini 3 Flash:

• Text to Text Safety: -3.9% (decrease in safety violations).
• Multilingual Safety: -2.6% (decrease in safety violations).
• Image to Text Safety: 0% (no change).
• Tone (objective tone of model refusal): +8.9% (improvement).
• Unjustified-refusals (model’s ability to respond to borderline prompts while remaining safe): +0.8% (non-egregious improvement).

Human red teaming, conducted by specialist teams, confirmed that Gemini 3.5 Flash met required launch thresholds for child safety. Across general content safety policies, including child safety, it exhibited similar or improved performance compared to Gemini 3 Flash. Red teaming also covered issues outside strict policies, comparing performance to Gemini 3.1 Pro, and found no egregious concerns.

A Frontier Safety Assessment concluded that Gemini 3.5 Flash, while excelling in agents and coding, does not introduce meaningful new capabilities or material increases in performance that would push it toward Critical Capability Levels (CCLs) compared to Gemini 3.1 Pro, which itself did not reach any CCLs. Additional testing in the cyber domain, prompted by previous Gemini 3 series models reaching the alert threshold, confirmed that Gemini 3.5 Flash remains below the cyber CCL. Risks and mitigations are further detailed in the Gemini 3.1 Pro model card.