Tech • IA • Crypto
Gemini 3.5 Flash: How to better use agentic AI?
8/10TL;DR
Gemini 3.5 Flash marks a shift toward agent-based AI workflows, forcing users to rethink costs, prompting methods, and how they manage AI as a team of digital workers.
KEY POINTS
Shift from Unlimited AI to Cost-Constrained Usage
The era of near-unlimited AI access is ending as subscription tiers and token-based pricing become standard. Users now face usage caps that can halt work for days, even on paid plans. This shift is pushing both individuals and companies to rethink how they allocate AI resources and control spending across tools like Gemini, ChatGPT, and Claude.
Rise of Agentic AI Systems
Gemini 3.5 Flash represents a move beyond traditional chatbots into “agentic” systems capable of making decisions, calling tools, and executing tasks autonomously. Instead of simple question-answering, these systems act more like digital workers, requiring users to take on managerial roles overseeing multiple AI agents.
New Subscription Economics
Google’s pricing structure has expanded with AI Pro, Ultra (€99/month), and Ultra premium (€219/month) tiers offering up to 20× higher limits. Additional costs for image and video generation credits further increase total spending. Enterprise plans, including discounted Workspace integrations, offer alternative access with business-oriented features.
Competitive Pricing Landscape
Gemini 3.5 Flash is positioned as a cost-efficient option at around $1.50 input / $9 output per million tokens, significantly cheaper than ChatGPT 5.5 ($5 / $30) and Claude Opus ($7.5 / $37.5). Competitors like MiniMax 3 are driving prices even lower, intensifying competition in the AI market.
Token Consumption and Workflow Optimization
Agentic systems consume tokens differently because they perform multi-step reasoning and tool usage. This increases costs unpredictably, making monitoring essential. Techniques such as tracking token usage, switching accounts, and managing sessions locally are emerging as practical workarounds.
Performance Strengths and Weaknesses
Gemini 3.5 Flash shows strong performance in coding and interactive outputs, including game generation and UI rendering. However, it struggles with complex context understanding, especially in tasks like infographic generation, where it may produce irrelevant or incorrect outputs compared to competitors like ChatGPT.
Prompt Engineering Becomes Critical
Traditional “loose” prompting is no longer effective. Users must design structured prompts with clear constraints, context blocks, and expected outputs. This includes defining rules, tools, validation criteria, and iteration logic, effectively turning prompts into deterministic programs rather than simple instructions.
End of “Voodoo Prompting”
Common techniques such as assigning vague roles (“you are an expert”) or asking AI to “think step by step” are increasingly unreliable. These approaches introduce ambiguity and variability, which is problematic in production environments where consistency and auditability are required.
Structured Prompts as Control Systems
Effective prompts now include defined inputs, constraints, decision rules, and output formats such as JSON. This allows AI systems to self-evaluate, retry tasks, and produce testable results. The goal is to reduce ambiguity and guide the model’s probabilistic behavior toward predictable outcomes.
AI as Managed Workforce
The evolving paradigm positions users as managers of AI teams rather than operators of single tools. Each agent can handle tasks like coding, research, or communication, but requires supervision, validation, and proper configuration to function reliably.
Enterprise and Automation Potential
Gemini 3.5 Flash performs strongly in enterprise scenarios such as RAG systems, Slack integrations, and automated workflows. Its speed and lower cost make it attractive for businesses deploying scalable AI agents, provided they implement proper governance and security controls.
CONCLUSION
The adoption of Gemini 3.5 Flash reflects a broader transformation toward agent-driven AI, where efficiency depends less on access and more on structured prompting, cost control, and the ability to manage AI systems as operational teams.