Google Gemini 3 Flash vs Google Gemini 2.0 Flash Thinking

Detailed comparison for LLMs

GoogleGoogle

On Guardion's LLM vulnerability Benchmark, Google Gemini 3 Flash is the more secure of the two: Gemini 3 Flash scores 10.0% and Gemini 2.0 Flash Thinking scores 21.5% on attack success rate (ASR) (lower is better). One or both scores are estimated from public safety evaluations pending a Guardion benchmark run.

Head-to-Head Overview

Gemini 3 Flash is the overall winner in this comparison!

Attack Success Rate (lower is safer)

ASR for Google Gemini 3 Flash vs Google Gemini 2.0 Flash Thinking. Green marks the safer model on each metric. Only the overall score is available for estimated models.

Overall (ASR)

Gemini 3 Flash
10.0%
Gemini 2.0 Flash Thinking
21.5%

TAP Attack Method (ASR)

Gemini 3 Flash
Gemini 2.0 Flash Thinking
100.0%

Crescendo Attack Method (ASR)

Gemini 3 Flash
Gemini 2.0 Flash Thinking
100.0%

Zero-Shot (ASR)

Gemini 3 Flash
Gemini 2.0 Flash Thinking
100.0%

Key Highlights

  • Google Gemini 3 Flash has a lower Overall (ASR).

Security Profile

Outward is better on every axis.

OverallTAPCrescendoZero-Shot
Gemini 3 Flash
Gemini 2.0 Flash Thinking
Full security profile
Google Gemini 3 Flash
Full security profile
Google Gemini 2.0 Flash Thinking

Frequently asked questions

Is Google Gemini 3 Flash or Google Gemini 2.0 Flash Thinking more secure?

On Guardion's LLM vulnerability Benchmark, Google Gemini 3 Flash is the more secure of the two: Gemini 3 Flash scores 10.0% and Gemini 2.0 Flash Thinking scores 21.5% on attack success rate (ASR) (lower is better). One or both scores are estimated from public safety evaluations pending a Guardion benchmark run.

What is the attack success rate (ASR) of Gemini 3 Flash vs Gemini 2.0 Flash Thinking?

Gemini 3 Flash has a 10.0% ASR and Gemini 2.0 Flash Thinking has a 21.5% ASR — the share of adversarial prompts that succeed across zero-shot, TAP, and Crescendo attacks. Lower is safer.

How were Gemini 3 Flash and Gemini 2.0 Flash Thinking tested?

Both were red-teamed with the HarmBench framework across zero-shot, TAP (Tree of Attacks with Pruning), and Crescendo multi-turn attacks, scored by Attack Success Rate.

Related Comparisons