Gemini Flash Gets Pricey, AI Act Delays, Agents Drive Online Traffic

2026-05-29 · Source: The Batch | DeepLearning.AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Legal & Regulatory · Depth: Advanced, long

Summary

Google launched Gemini 3.5 Flash, a faster, mid-tier multimodal model offering improved agentic capabilities and visual understanding. It supports diverse inputs up to 1 million tokens and outputs up to 64,000 tokens at 204 tokens/second. While it excels in benchmarks like APEX-Agents-AA and MMMU-Pro, its price is three times that of Gemini 3 Flash, reflecting a trend of rising per-token costs. Concurrently, the EU amended its AI Act, delaying "high-risk" system restrictions until December 2027 and easing burdens for smaller companies, driven by competitiveness concerns. A Human Security report revealed AI-driven internet traffic nearly tripled in 2025, with AI agents showing 80x growth, primarily on product and search pages, alongside a 47% rise in malicious scraping. Finally, researchers introduced a fine-tuning method for text-to-image generators that uses a staged "plan, sketch, inspect, refine" process, improving spatial relationship accuracy and object attributes, boosting BAGEL-7B's GenEval score from 77% to 83%. The article also notes the rise of AI Forward Deployed Engineers but predicts a larger demand for generalist AI Engineers.

Key takeaway

For Directors of AI/ML evaluating team structures, prioritize hiring your own generalist AI Engineers. This maintains vendor optionality and fosters internal expertise, rather than relying on vendor-embedded FDEs. New models like Gemini 3.5 Flash offer performance gains but incur significantly higher per-token costs; plan your budget carefully. Prepare your infrastructure for rapidly increasing AI-driven internet traffic. Also, consider advanced techniques like staged image generation to improve your model output quality.

Key insights

AI is creating new roles, driving up model costs, reshaping regulations, and advancing agentic capabilities and image generation.

Principles

• Step-by-step reasoning improves AI output quality.
• Overregulation can stifle innovation and competitiveness.
• Companies prefer internal AI talent for vendor optionality.

Method

Fine-tune a multimodal model to compose images by cycling through "plan, sketch, inspect, refine" stages, using GPT-4o to generate training data for each step.

In practice

• Implement staged image generation for complex spatial prompts.
• Monitor internet traffic for new AI agent patterns.
• Prioritize internal AI engineering talent for flexibility.

Topics

• AI Engineering Roles
• Gemini 3.5 Flash
• Multimodal AI Models
• AI Regulation
• AI Agent Traffic
• Text-to-Image Generation

Best for: CTO, VP of Engineering/Data, Executive, AI Engineer, Director of AI/ML, AI Scientist

Related on AIssential

• Google's Gemini 3.5 Flash follows Anthropic and OpenAI in making newer AI models significantly pricier — Newer, more capable AI models like Gemini 3.5 Flash are significantly pricier due to increased token consumption for complex agentic tasks.
• 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.
• Gemini 3.5 Flash might be fast enough for gen AI to make sense — Gemini 3.5 Flash demonstrates that high intelligence and efficiency can enable scalable agentic AI.
• The Model Eats the Scaffolding: DeepMind's Logan Kilpatrick & Tulsee Doshi on 3.5 Flash, Omni & More — Google prioritizes cost-effective, fast AI models and integrated agentic harnesses for broad product integration and user experience.
• Google just declared itself a contender in AI design — Google's "AI-first" strategy is delivering multimodal, agentic AI across its product suite, from design to commerce and scientific research.
• Google launches Gemini 3.5 Flash as its default AI model — Google's Gemini 3.5 Flash offers high-speed, cost-effective AI for agentic tasks, becoming the default model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Batch | DeepLearning.AI.