왜 provider가 달라 — 마케팅 너머

같은 task, 다른 shape

다른 provider가 같은 API 뒤에 다른 모델만은 아니야. tokenizer, role 이름, structured-output 메커니즘, tool-call format, system-prompt semantics, refusal calibration, reasoning configuration, multimodal capability, pricing, rate limit, SDK convention 다 달라. 차이 add up; 한 곳에 tuned된 prompt이 다른 곳에 cleanly land rarely.

2026에 알 4개 family

Anthropic — Claude family. Long context, XML-tagged prompt, extended thinking, tool use 강해.
OpenAI — GPT-5.5 / o-series. Structured output (strict JSON schema), reasoning, multimodal 강해.
Google — Gemini 2.5 Pro / Flash. Long context, multimodal, native search grounding 강해.
Open-source — Llama 4, Qwen 3, Mistral 등. 어디든 run; quality 다양; provider lock-in 없음.

왜 중요한가

provider 사이로 이동할 거야 — cost, outage hedging, regional compliance, capability 위해. plan해. Prompt-engineering work는 portable; wiring 아니야.

Code

한 표로 차이 (mental model)·plaintext

                  Claude         OpenAI         Gemini         OSS
roles             system+user    system+dev+u   system+user    varies
XML tag culture   strong         weak           moderate       varies
strict JSON       via tools      json_schema    response_schema partial
reasoning         extended       o-series       thinking        partial
tool_choice       any/specific   required/named auto/any        varies
token counter     server         tiktoken       SDK            varies

왜 provider가 달라 — 마케팅 너머

같은 task, 다른 shape

2026에 알 4개 family

왜 중요한가

Code

External links

Exercise

Progress

댓글 0