o3, o4-mini, and reasoning models: when paying for thinking time is worth it

o3 and o4-mini are OpenAI's reasoning branch. They work differently from regular chat models and should be used for different tasks.

o3 and o4-mini are not in the Yenisei catalog, but open-weight reasoning models are: Kimi K2 Thinking (kimi-k2-thinking), DeepSeek V4 Pro (deepseek-pro), and GPT-OSS 120B and GPT-OSS 20B, OpenAI's open models with reasoning.

How a reasoning model differs from a regular one

A regular model answers almost immediately: it produces the first coherent solution. A reasoning model "thinks" first: it spends compute on an internal chain of reasoning, checks options, discards dead ends, and only then answers.

You pay for this in time and money, and you gain accuracy on tasks where the first plausible answer is often wrong.

At OpenAI, the branch's top model, o3, handles math, proofs, and multi-step planning, while the light o4-mini takes simpler tasks, faster and cheaper. The same "large plus light" pair in Yenisei is Kimi K2 Thinking and GPT-OSS 20B (gpt-oss-20b).

Where reasoning wins

• Tasks with a clear correctness check: math, algorithms, untangling complex conditional logic.
• Planning where many constraints have to be held at once.
• Debugging where the symptom is far from the cause and hypotheses need to be worked through.

Where reasoning is a waste

• Simple text generation, rewriting, styling: there is nothing to think about, and you overpay for latency.
• A high-volume stream of short requests: a reasoning model responds more slowly and the queue grows.
• Creative tasks: the extra "discipline" sometimes gets in the way.

Example call

from openai import OpenAI
client = OpenAI(base_url="https://api.yenisei.ru/v1", api_key=YENISEI_KEY)
resp = client.chat.completions.create(
    model="kimi-k2-thinking",
    messages=[{
        "role": "user",
        "content": "There are 12 coins, one of them counterfeit, and a balance scale with no weights. In 3 weighings, find the counterfeit coin and say whether it is lighter or heavier.",
    }],
)
print(resp.choices[0].message.content)

How to choose

Start with the light one, GPT-OSS 20B: it is cheaper and handles most logic tasks. If a task is beyond it (a complex proof, a long planning chain), move up to Kimi K2 Thinking or DeepSeek V4 Pro.

The general rule: use reasoning models selectively, for a specific hard task, not as a replacement for a chat model. For a stream of conversations, light models remain the choice: Mistral Nemo (mistral-nemo) or Yenisei 100 (yenisei-100). Payment is in rubles, 1 credit = 1 ₽.