Gemini 3.5 Flash for stock trading

Fast and cheap with a million tokens of context is an unusual combination, and it is the reason to look at Gemini 3.5 Flash. Most models make you pick. This one lets you run a tight scan loop and, when you need it, hand the same model a pile of history without switching providers or building a retrieval layer first.

Use the big window deliberately, not reflexively. You pay for input tokens on every call, so a loop that runs constantly with a bloated prompt turns a cheap model into an expensive one. Recall also degrades in the middle of very long prompts, which means burying the one number that matters between two hundred thousand tokens of filler is a way to get it ignored. The pattern that works: a small prompt for the loop, the large window reserved for a periodic review where you actually want every trade and every written rationale in front of the model at once. If your prompt prefix is stable, look at Google's context caching before you optimize anything else, and check current pricing rather than trusting a number from a blog post.

Structured output is a strength. The API supports schema-constrained responses, so instead of asking politely for JSON and hoping, you declare the shape and get it. For a trading agent that removes an entire category of parsing bug. Tool calling is dependable, with one caveat worth planning for: Flash will happily issue parallel or repeated tool calls, and if your order endpoint is not idempotent you can end up submitting the same trade twice. Give every order a client-side identifier and deduplicate server side. That is cheap to build and painful to retrofit.

The weaknesses are the usual Flash weaknesses. It hedges, it pads, and it will produce three qualifications where a peer would produce a verdict. Prompt for a decision and a reason and cap the length. On genuinely hard analysis it trails Pro by enough to notice. Safety filtering also trips occasionally on financial phrasing, which is configurable but is the sort of thing that surfaces at three in the morning if you never tested for it. Build the loop on Flash, run the review on something heavier.