Choosing a vendor and defending its flag for the rest of the week is work I can do without. On July 11, 2026, I suggested trying Gemini for frontend work and assessed Claude as good for backend work. I was thinking about distributing work by function. The Gemini suggestion was still a proposed experiment.
On July 21, Google announced Gemini 3.6 Flash and 3.5 Flash-Lite, with different propositions and prices. More options make that distribution interesting, but they also require a choice. A new name enters the conversation; the task remains the reference for deciding where it fits and what I need to verify before accepting its output.
For frontend work, I want to examine what someone uses and how the interface behaves. Backend work requires attention to other parts of the system. Trying different tools against those requirements makes sense. My opinion of Claude can guide candidate selection, and my interest in Gemini can guide an experiment. Neither settles the decision for every subsequent request.
Moving between tools costs work too. If I have to reconstruct the request at each switch, a promising division quickly becomes tedious. I want to preserve the expected result and settled decisions in a form the next executor understands. Retelling the entire conversation turns model variety into another administrative task for me.
The boundary between frontend and backend needs an owner. In a hypothetical example, a screen expects a response that another part of the system is still defining. Sending each side to a different model does not resolve that dependency. Somebody needs to follow whether the behavior prepared on one side supports what the other is proposing.
Flash-Lite adds the question of how much model an activity needs. A simple task and a long investigation may justify different choices. I want to discover when additional capability helps delivery and when another option can handle the work. The number of names in the selector should not become an obligation to pick the most impressive one every time.
I will keep distribution by function as a working hypothesis and decide who follows dependencies before splitting a task that crosses both sides. Claude remains an option I assessed well for backend work; Gemini for frontend deserves the proposed experiment. The new Flash models enter as candidates to evaluate on suitable requests. My commitment is to a delivery whose parts work together, even when each part comes from a different tool.