Claude vs GPT vs Gemini vs DeepSeek: which model for which job

Why there is no single best model

Each provider ships several sizes, from large flagship models to small fast ones, and updates them every few months. Public benchmarks measure narrow tasks and go out of date quickly. What matters is how a model handles your prompts, your language, your code base and your budget. With one key for all of them, testing is cheap: you change the model parameter and run the same request again.

This guide describes well-known general traits, not benchmark scores. Treat it as a starting shortlist, then test.

Quick comparison

Which exact versions are available, and their names, are listed on the models page. Always copy the model name from there.

Best model for coding

For working inside a code base, editing several files and running tools, Claude models are a common default; Claude Code is built around them. GPT models are also widely used for coding and are strong at following a strict output format. DeepSeek is a good choice when you need lots of code explanations, tests or refactors and cost matters more than the last bit of quality.

• Agent work (multi-file edits, running commands): start with hk-5-5.
• Quick questions, small functions, code review comments: try a smaller model or deepseek-v4.1-flash.
• Generating many unit tests or docstrings in bulk: compare deepseek-v4.1-flash with a small GPT or Gemini model on cost.

Best model for writing

For marketing copy, documentation and emails, Claude is often preferred for natural tone and for sticking to a style guide; GPT is a strong all-rounder and good at short punchy variants. For writing in languages other than English, test each model in that language, since quality varies by language more than by provider.

Best model for long documents

For contracts, reports, transcripts or large code dumps, the context window (how much text fits in one request) matters first. Gemini models are known for very large context windows, and Claude models also handle long inputs well. Remember that every token you send is billed as input, so a long document costs input tokens on every request. For repeated questions about the same document, consider sending only the relevant sections.

Best model for cheap bulk work

Classification, tagging, extraction, translation of short strings and summarizing many small items are jobs where a small, cheap model is usually enough. Candidates are DeepSeek and the small or Flash-type variants of Gemini, GPT and Claude. Compare the prices per 1 million tokens:

Real prices now

Estimated from the prices on the models page: 2,000 input tokens and 500 output tokens per chat. The console shows the actual charge.

For a bulk job, the output price often matters as much as the input price. Ask for short answers (a label, a JSON object) and set max_tokens so a model cannot write an essay when you need one word. See how token pricing works.

How to test models on your own prompts

1. Collect 10 to 20 real prompts from your app or workflow, including a few hard ones.
2. Pick two to four candidate models from the models page.
3. Run every prompt through every model with the script below.
4. Score the answers for what matters to you: correctness, tone, format, length.
5. Compare the token counts and the cost per call on Usage.
6. Choose the cheapest model that passes your quality bar, and re-test when new versions appear.

import os, json
from openai import OpenAI

client = OpenAI(base_url="https://api.macloudai.com/v1", api_key=os.environ["MACLOUD_API_KEY"])
models = ["hk-5-5", "tt-6-luna", "gem-3.8-flash", "deepseek-v4.1-flash"]
prompts = [
    "Extract the invoice number and total from: 'Invoice INV-2291, total due 1,250.00'. Reply as JSON.",
    "Rewrite this sentence to sound friendly: 'Your request has been rejected.'",
]

for p in prompts:
    for m in models:
        r = client.chat.completions.create(
            model=m,
            messages=[{"role": "user", "content": p}],
            max_tokens=300,
        )
        print(json.dumps({
            "model": m,
            "in": r.usage.prompt_tokens,
            "out": r.usage.completion_tokens,
            "answer": r.choices[0].message.content,
        }, ensure_ascii=False))

You can also compare models by hand in the web chat at /chat, which uses the same balance. In the console, Usage shows the cost of each test call next to the model name.

Mixing models in one app

Because every model uses the same endpoint and key, you do not have to pick just one. A common pattern is a cheap model for routing, classification or first drafts and a stronger model only for the hard cases. Another is a fallback: if one model returns a 5xx error, retry the same request with a different model name.

FAQ

More guides

• One API key for Claude, GPT, Gemini and DeepSeek
• How to use Claude Code with a custom API endpoint
• Claude Code API errors and fixes: 401, 403, 404, 429, 5xx
• How to use Cursor with your own API key and base URL
• How AI API token pricing works and how to estimate your bill
• Add AI to n8n workflows with an OpenAI-compatible API