I would like to share my early impressions of Gemini 3.5 Flash after using it extensively in real-world AI agent workflows.
From a task execution perspective, Gemini 3.5 Flash is highly capable and often performs better than previous Gemini models when handling structured tasks. However, one issue I consistently encounter is token consumption. In my testing, Gemini 3.5 Flash appears to consume tokens at a higher rate than Gemini 3.1 Pro, which can significantly impact operational costs when running long agent sessions.
I would also like to suggest adding Claude Sonnet 4 as an available model option instead of focusing primarily on Claude Sonnet 4.6 Thinking variants.
In my experience, reasoning-heavy “thinking” models are not always the optimal choice for agent-based workflows. When building agents that must handle multiple responsibilities such as writing specifications, planning, coding, reviewing code, and executing tasks, model specialization often produces better results than relying on a single model for everything.
Personally, I use different models for different purposes:
-
Gemini models for fast iteration and implementation.
-
Claude models for structured reasoning and documentation.
-
GPT models for broader problem-solving and general-purpose assistance.
Each model has strengths and weaknesses, and I believe the most effective workflow comes from leveraging those strengths appropriately.
One reason I appreciate Gemini 3 Flash is its ability to accelerate development cycles. I typically review and audit generated code before allowing the agent to continue, and the model performs well in that workflow.
Regarding Claude Sonnet 4.6 and Claude Opus 4.6, my concern is not primarily the cost. The larger issue is response latency and token utilization. After extensive testing across GPT, Claude, and Gemini families, I have found that additional reasoning time and token consumption do not always translate into proportional productivity gains.
I also tested Gemini 3.5 Flash using both Medium and Low settings to optimize costs. Even under Low settings, token usage remained relatively high compared to the value delivered for my specific use cases.
For this reason, I would recommend adding:
-
Gemini 3 Flash
-
Claude Sonnet 4
Claude Sonnet 4 offers what I consider an excellent balance between capability, speed, and cost. For agent-based coding workflows that follow predefined plans and specifications, it delivers fast responses while remaining significantly more cost-effective than larger reasoning-focused models.
Finally, I would like to encourage the community to avoid becoming overly attached to any single AI model.
No model is universally best. Every model excels in different areas. By assigning the right tasks to the right models, developers can reduce audit time, minimize refactoring effort, decrease repetitive debugging cycles, and improve overall productivity.
The future of AI-assisted development is not about finding one perfect model-it is about building efficient workflows that leverage the strengths of multiple models working together.