You're prompting GPT-5.5 for English-to-Spanish translation. The zero-shot version works fine. When would adding few-shot examples reliably outperform the zero-shot prompt?

Zero-shot prompting just describes the task: translate this sentence from English to Spanish. Few-shot prompting adds a small number of input-output example pairs before the actual request. The intuition used to be that more examples are always better, but that is no longer true with modern frontier models. For common tasks like translation, summarization, or sentiment classification, models like Claude Opus 4.7 and GPT-5.5 already know how to do the task. They were trained on enormous amounts of multilingual text. Showing them three more translations does not teach them anything new about Spanish. Where few-shot still wins is when you need something specific that the model cannot guess. A particular output format like JSON with named fields, a niche style like legal phrasing, or a domain pattern like a company-specific summary structure. The examples lock in the shape, not the knowledge. The trap is adding examples by reflex. They cost tokens, add latency, and if the examples are noisy, they can actively make the output worse.

Few-shot examples teach format and style, not knowledge. That is the right frame for choosing between zero-shot and few-shot in 2026. For common tasks on capable models, zero-shot is already strong. Translation, summarization, classification, basic reasoning: Claude Opus 4.7 and GPT-5.5 hit ceiling-near accuracy from a clean instruction alone. Adding examples may help by a point or two, or may hurt if the examples are inconsistent. The cost is more input tokens and slightly higher latency. Few-shot reliably wins in three cases. First, format adherence: you need a specific output schema, layout, or structured response that the zero-shot prompt occasionally violates. Second, niche tasks: a domain-specific style, an unusual labeling scheme, a company-internal terminology the model has not seen. Third, structured outputs that have to match a downstream parser exactly. The trap is treating few-shot as a generic accuracy upgrade. It is not. Bad or noisy examples actively mislead the model, and good examples cost tokens. The senior move is to A/B test zero-shot vs few-shot on a golden set rather than assume the few-shot version wins.