How to choose a model for the task: a practical checklist

There are dozens of models in the catalog, and "take the flagship for everything" is a bad strategy: overpayment and needless latency. The choice boils down to four questions. Go through them in order and you will be left with one or two candidates.

1. What kind of task

This rules out most options right away:

• Classification, field extraction, short answers: Flash/Mini models. Cheap, fast, good enough.
• Complex logic, math, planning, debugging: reasoning models (kimi-k2-thinking, gpt-oss-120b) or a strong Pro (deepseek-pro).
• Code: specialized models (qwen3-coder-next, kimi-k2.7-code, yenisei-code).
• Texts and content: Pro for what matters, Flash for volume.
• Images, audio: models that accept images or sound (qwen-flash, mimo-v2.5, voxtral-small).
• Knowledge base search: embeddings (bge-m3) plus any chat model for generation.

2. How much context you need

Estimate the maximum input: dialog history, documents, code.

• Up to about 8K: almost any model will do.
• Tens of thousands of tokens (long dialogs, RAG): take a model with a context of 128K or more.
• Whole code modules or large documents at once: Kimi and other long-context models.

Do not chase the maximum "just in case": a large context also means a large bill for input tokens.

3. Speed over brains, or the other way around

• Interactive (chat, autocomplete, voice): latency is critical. Flash plus streaming.
• Background processing (overnight document analysis, reports): a slow flagship is fine, nobody is waiting.

The rule: if a person is watching the cursor and waiting, speed. If the result goes into a queue, quality.

4. What it will cost

Estimate the volume: requests per day times tokens per request. At scale, the 5 to 20x Flash/Pro difference turns into thousands of rubles. 1 credit = 1 ₽, so count straight in rubles. For rare calls (a couple hundred a day) you can almost ignore the price; for thousands, it decides.

Check the Russian

A separate item for a Russian-language product: not all models write Russian equally well. DeepSeek, Qwen, Kimi write cleanly, without calques. Run a couple of typical requests from your domain before choosing.

Don't choose blindly: test

The checklist narrows the choice to two or three candidates. From there, do not argue in chat; run both on 20 to 30 real examples of your task with reference answers:

for model in ["qwen-flash", "deepseek-flash"]:
    for case in test_set:
        r = client.chat.completions.create(
            model=model, messages=case["messages"], temperature=0)
        score(model, case, r)  # compare with the reference, log usage

Look at quality and usage at the same time. Often the cheap model loses to the flagship by 2 to 3% accuracy at a tenth of the price, and for your task that is a good trade.

A short cheat sheet

• Mass inference: Flash.
• Complex and important: Pro / reasoning.
• Code: coder models.
• Long input: long-context.
• Multimedia: models for images and sound.
• Search over data: embeddings.

Start with the cheapest model that passes your quality test, and move up only where it genuinely falls short. That is cheaper than starting with the flagship and then looking for where to cut.