Claude Haiku 4.5 and fast models for high-volume requests: what to use in Yenisei

Claude Haiku 4.5 is the lightest model in Anthropic's lineup. The larger models are used for long, complex tasks; Haiku is for short requests that come in large numbers.

There is no Claude in the Yenisei catalog. But "fast and cheap at volume" is one of the most common needs, and we have our own models for it.

What people value in Haiku

• Speed: noticeable in chatbots and wherever the answer goes to a person in real time.
• Price: suitable for bulk processing.
• 200K context: enough for most "hot" tasks.
• Reasoning: unusual for a light model; it doesn't fall apart on moderately hard tasks.

What to use in Yenisei

• Mistral Nemo (mistral-nemo): very cheap and fast, with tool calling. For classification and field extraction.
• GPT-OSS 20B (gpt-oss-20b): OpenAI's open model with reasoning at a low price, for when the stream isn't entirely trivial.
• DeepSeek V4 Flash (deepseek-flash): a fast model with reasoning and a long context.
• Yenisei 100 (yenisei-100): our light model, which also understands images.

Where to put a fast model

• Classifying tickets and routing requests.
• Extracting fields from text (dates, amounts, addresses).
• Short summaries before handing off to a larger model.
• Bot answers to frequent questions, where latency matters.

Example: classification in a stream

from openai import OpenAI
client = OpenAI(base_url="https://api.yenisei.ru/v1", api_key=YENISEI_KEY)
resp = client.chat.completions.create(
    model="mistral-nemo",
    messages=[
        {"role": "system", "content": "Return one category: payment, delivery, refund, other."},
        {"role": "user", "content": "I was charged twice for the same order."},
    ],
)
print(resp.choices[0].message.content)  # payment

Where a fast model doesn't belong

Don't use it for tasks where the cost of a mistake is high: complex refactoring, legal analysis, multi-step agents with branching. Those need a larger model, such as DeepSeek V4 Pro (deepseek-pro). A light model is good precisely because it handles a volume of repetitive work cheaply and quickly. Payment is in rubles: 1 credit = 1 ₽.