Gemini 3.5 FLASH: BAD to OUTSTANDING

2026-05-20 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

Google's new Gemini 3.5 Flash model, released May 19, 2026, shows expected performance relative to Gemini 3.1 Pro Preview, scoring 55.3 in general evaluations compared to Pro's 57.2. For logic and causal reasoning, Flash achieved 40% and 72% success rates, slightly below Pro's 44% and 77%. Critically, Flash is priced significantly lower, at \$1.5 per million input tokens and \$9 per million output tokens, half the cost of Pro for high-volume usage. A live causal reasoning test, involving a complex elevator puzzle, revealed high solution variance. Initially, Flash produced poor solutions (e.g., 16 button presses), but after 5 minutes 27 seconds and multiple optimization attempts, it achieved the theoretically optimal 7-button solution plus an emergency exit. However, it initially failed to output numerical steps, requiring a corrective prompt.

Key takeaway

For Machine Learning Engineers evaluating LLMs for complex, cost-sensitive reasoning tasks, Gemini 3.5 Flash presents a compelling option due to its significantly lower price point. However, you must account for its high solution variance and potential need for iterative prompting to achieve optimal results. Plan to monitor reasoning traces and be prepared to issue corrective prompts, as initial outputs may lack critical details or be suboptimal.

Key insights

Gemini 3.5 Flash offers significant cost savings but exhibits high solution variance and requires iterative prompting for complex tasks.

Principles

• Model performance can show high variance on complex, multi-constraint problems.
• Lower-tier models can achieve optimal solutions with sufficient compute/time.

In practice

• Consider Gemini 3.5 Flash for cost-sensitive applications.
• Anticipate iterative prompting and longer processing times for optimal results.

Topics

• Gemini 3.5 Flash
• Large Language Models
• Model Benchmarking
• Causal Reasoning
• AI Model Pricing
• Performance Variance

Best for: AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Gemini 3.5 Flash Looks Good For How Fast It Is — Gemini 3.5 Flash offers high speed for agentic tasks, but its intelligence and cost-effectiveness are debated.
• Gemini 3.1 Flash-Lite — Gemini 3.1 Flash-Lite offers a cost-effective alternative with adjustable thinking levels.
• Gemini 3.5 Flash: more expensive, but Google plan to use it for everything — Gemini 3.5 Flash signals a new phase of higher LLM pricing and broad internal deployment across Google's ecosystem.
• DeepSeek FLASH Destroys Gemini FLASH — DeepSeek v4 Flash consistently outperforms Gemini 3.1 Flash in problem-solving efficiency and transparency.
• Gemini 3.5 Flash Test | Coding, OCR, Image Understanding, Pricing, Speed | 🔴 Live — Gemini 3.5 Flash offers significant speed improvements and strong document understanding, but its coding and frontend capabilities are inconsistent.
• TAI #206: Gemini 3.5 Flash Is Stronger, But Flash Is No Longer Cheap — Google's latest Gemini models offer strong capabilities but come with significantly increased costs and deployment challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.