Gemini 3.1 Pro 對 Gemini 3 Pro：故意變慢更聰明

🔑 增強重點摘要

• •Gemini 3.1 Pro achieves a verified 77.1% score on ARC-AGI-2 benchmark, more than doubling Gemini 3 Pro's 31.1-35% performance in abstract reasoning[2][4][6][7][9].
• •Introduces a 3-level thinking system (low, medium, high with Deep Think Mini), where medium level matches 3 Pro's high reasoning but with lower latency[4].
• •Offers 15% quality improvement with fewer output tokens and fixes response truncation issues reported in Gemini 3 Pro[4][7].
• •Adds native SVG handling for precise code writing and animation, plus highest score on GPQA Diamond graduate-level science benchmark[2][7].

🛠️ 技術深入

• •3-level thinking system: low (minimal inference, fast), medium (balanced, ≈ Gemini 3 Pro high but lower latency), high (Deep Think Mini, deepest inference)[4].
• •Context window: 1 million tokens (both Gemini 3.1 Pro Preview and Gemini 3 Pro)[3][5].
• •Enhanced safety and reliability: outperforms Gemini 3 Pro in safety evaluations and long-task stability while keeping unjustified refusals low[4][8].
• •Output efficiency: 15% quality boost using fewer tokens, resolves truncation in long responses[4][7].
• •Native multimodal capabilities including SVG vector graphics processing and animation[2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線