Gemini 3.5 Flash

Google DeepMind / Gemini · 2026-05-19 · 通用模型

Gemini 3.5 Flash

Google 在 I/O 2026 发布 Gemini 3.5 首个模型 Gemini 3.5 Flash，定位「frontier intelligence with action」，并成为 Gemini 应用与 Search AI Mode 的默认模型。评测横跨终端智能体、多模态理解与抽象推理，亮点如 Terminal-Bench 2.1 76.2%、MCP Atlas 83.6%，页面并称输出速度约为其他前沿模型 4 倍。

输入模态: 文本 / 图像
上下文: 1M
参数: 官方资料未说明
价格（每百万 tokens）: 无此口径报价；不等于免费

本变体的评测证据

terminalbench 76.2% 模型 gemini-3-5-flash · 版本 2.1 · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-08-31 · 距快照 34 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding · quote_snippet: outperforming Gemini 3.1 Pro on challenging coding and agentic benchmarks like Terminal-Bench 2.1 (76.2%)

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Maps to existing benchmark terminalbench, variant 2.1. Comparison baseline named as Gemini 3.1 Pro (that 3.1 release has no library entry yet — inventory gap flagged in README).

打开官方来源

gdpval-aa 1656 Elo 模型 gemini-3-5-flash · 版本未说明 · 指标 knowledge_work_elo · 单位 elo 来源等级 A · vendor_reported · 核对日期 2026-08-31 · 距快照 34 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding · quote_snippet: GDPval-AA (1656 Elo)

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": "leaderboard-style Elo",
  "judge": null
}

Reuses candidate id gdpval-aa (registered batch 1). Page writes 'GDPval-AA' without the v2 suffix used on later Google pages (3.6/3.7 write GDPVal-AA v2) — variant left null here; do not assume same edition as v2 rows without checking.

打开官方来源

mcp-atlas 83.6% 模型 gemini-3-5-flash · 版本未说明 · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-08-31 · 距快照 34 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding · quote_snippet: MCP Atlas (83.6%)

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Reuses candidate id mcp-atlas (registered batch 1). Muse Glimmer model card later reports 'MCP Atlas (Public)' 75.5 — the Public-subset qualifier matters for comparability.

打开官方来源

charxiv-reasoning 84.2% 模型 gemini-3-5-flash · 版本未说明 · 指标 accuracy · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-08-31 · 距快照 34 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding · quote_snippet: leading in multimodal understanding (84.2% on CharXiv Reasoning)

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "time_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Reuses candidate id charxiv-reasoning (registered batch 2). The Gemini 3.6 model page later splits CharXiv into no-tools 84.2 / with-tools 84.9 for 3.5 Flash — this row's condition (likely no-tools) is not stated on this page.

打开官方来源

swebench-pro 55.1% (Public, single attempt) 模型 gemini-3-5-flash · 版本 Public, single attempt · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: SWE-Bench Pro (Public) — Diverse agentic coding tasks, single attempt · figure: gemini-3-5__benchmarks__light.gif — SWE-Bench Pro (Public) — Diverse agentic coding tasks, single attempt row, Gemini 3.5 Flash column · quote_snippet: SWE-Bench Pro (Public): single attempt

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 49.6% / Gemini 3.1 Pro 54.2% / Claude Opus 4.7 64.3% (best in row) / GPT-5.5 58.6%; Sonnet 4.6 em-dash.

打开官方来源

toolathlon 56.5% 模型 gemini-3-5-flash · 版本未说明 · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: Toolathlon — Real-world general tool use · figure: gemini-3-5__benchmarks__light.gif — Toolathlon — Real-world general tool use row, Gemini 3.5 Flash column · quote_snippet: Toolathlon: real-world general tool use

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 49.4% / GPT-5.5 55.6%; 3.1 Pro, Sonnet 4.6 and Opus 4.7 em-dash. Gemini 3.5 Flash best in row.

打开官方来源

osworld 78.4% 模型 gemini-3-5-flash · 版本 Verified · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: OSWorld-Verified — Agentic computer use · figure: gemini-3-5__benchmarks__light.gif — OSWorld-Verified — Agentic computer use row, Gemini 3.5 Flash column · quote_snippet: OSWorld-Verified: agentic computer use

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 65.1% / Gemini 3.1 Pro 76.2% / Claude Sonnet 4.6 72.5% / Claude Opus 4.7 78.0% / GPT-5.5 78.7% (best in row).

打开官方来源

finance-agent 57.9% 模型 gemini-3-5-flash · 版本 v2 · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: Finance Agent v2 — Financial analysis and decision-making · figure: gemini-3-5__benchmarks__light.gif — Finance Agent v2 — Financial analysis and decision-making row, Gemini 3.5 Flash column · quote_snippet: Finance Agent v2: financial analysis and decision-making

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 42.6% / Gemini 3.1 Pro 43.0% / Claude Sonnet 4.6 51.0% / Claude Opus 4.7 51.5% / GPT-5.5 51.8%. Gemini 3.5 Flash best in row.

打开官方来源

mmmu 83.6% (no tools) 模型 gemini-3-5-flash · 版本 Pro, no tools · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: MMMU-Pro — Multimodal understanding and reasoning, no tools · figure: gemini-3-5__benchmarks__light.gif — MMMU-Pro — Multimodal understanding and reasoning, no tools row, Gemini 3.5 Flash column · quote_snippet: MMMU-Pro: multimodal understanding and reasoning

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 81.2% / Gemini 3.1 Pro 80.5% / Claude Sonnet 4.6 74.5% / Claude Opus 4.7 75.2% / GPT-5.5 81.2%. Gemini 3.5 Flash best in row.

打开官方来源

blueprint-bench-2 33.6% (normalized score) 模型 gemini-3-5-flash · 版本 normalized score · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: Blueprint-Bench 2 — Agentic spatial reasoning, normalized score · figure: gemini-3-5__benchmarks__light.gif — Blueprint-Bench 2 — Agentic spatial reasoning, normalized score row, Gemini 3.5 Flash column · quote_snippet: Blueprint-Bench 2: agentic spatial reasoning

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 0.0% / Gemini 3.1 Pro 26.5% / Claude Sonnet 4.6 6.7% / Claude Opus 4.7 24.5% / GPT-5.5 36.2% (best in row).

打开官方来源

mrcr 77.3% (v2 8-needle, 128k average) 模型 gemini-3-5-flash · 版本 v2 (8-needle), 128k (average) · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: MRCR v2 (8-needle) — 128k (average) · figure: gemini-3-5__benchmarks__light.gif — MRCR v2 (8-needle) — 128k (average) row, Gemini 3.5 Flash column · quote_snippet: MRCR v2 (8-needle): long context performance, 128k average

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 67.2% / Gemini 3.1 Pro 84.9% / Claude Sonnet 4.6 84.9% / Claude Opus 4.7 59.3% / GPT-5.5 94.8% (best in row).

打开官方来源

mrcr 26.6% (v2 8-needle, 1M pointwise) 模型 gemini-3-5-flash · 版本 v2 (8-needle), 1M (pointwise) · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: MRCR v2 (8-needle) — 1M (pointwise) · figure: gemini-3-5__benchmarks__light.gif — MRCR v2 (8-needle) — 1M (pointwise) row, Gemini 3.5 Flash column · quote_snippet: MRCR v2 (8-needle): long context performance, 1M pointwise

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 22.1% / Gemini 3.1 Pro 26.3%; Sonnet 4.6, Opus 4.7 and GPT-5.5 em-dash (no 1M context). Separate row from the 128k average because the aggregation differs.

打开官方来源

hlehle 40.2% (full set, text + MM) 模型 gemini-3-5-flash · 版本 full set (text + MM), no tools · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: Humanity's Last Exam — Academic reasoning, full set (text + MM) · figure: gemini-3-5__benchmarks__light.gif — Humanity's Last Exam — Academic reasoning, full set (text + MM) row, Gemini 3.5 Flash column · quote_snippet: Humanity's Last Exam: academic reasoning, full set

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": null
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 33.7% / Gemini 3.1 Pro 44.4% / Claude Sonnet 4.6 33.2% / Claude Opus 4.7 46.9% (best in row) / GPT-5.5 41.4%.

打开官方来源

arc-agi 72.1% 模型 gemini-3-5-flash · 版本 2 (ARC Prize Verified) · 指标未说明 · 单位 percent 来源等级 A · vendor_reported · 核对日期 2026-09-01 · 距快照 33 天

归一化读数：不计算；样本及方差不完整，无法计算置信区间。

原文位置与完整协议

heading: 3.5 Flash: frontier performance for agents and coding (benchmark table) · row: ARC-AGI-2 — Abstract reasoning puzzles · figure: gemini-3-5__benchmarks__light.gif — ARC-AGI-2 — Abstract reasoning puzzles row, Gemini 3.5 Flash column · quote_snippet: ARC-AGI-2: abstract reasoning puzzles

{
  "harness": null,
  "tools": null,
  "shots": null,
  "reasoning_effort": null,
  "temperature": null,
  "top_p": null,
  "token_budget": null,
  "turn_limit": null,
  "run_count": null,
  "aggregation": null,
  "judge": "ARC Prize Verified"
}

Verified visually 2026-09-01 by reading the archived GIF (models/2026-05-19-gemini-3-5-flash/images/03.gif): Gemini 3 Flash 33.6% / Gemini 3.1 Pro 77.1% / Claude Sonnet 4.6 58.3% / Claude Opus 4.7 75.8% / GPT-5.5 84.6% (best in row).

打开官方来源