Choosing the Right LLM for the Job: Fast, Cheap or Smart

How to pick a language model that fits your real work, using six criteria and a test set built from your own tasks instead of leaderboards alone.

"Which model should we use?" may be the question IT teams hear most this year. Marketing wants one that writes beautiful Thai. Customer service wants fast answers. Finance wants it cheap. And leadership has just read about a new model that is supposedly the smartest yet. The truth is that no model is best at everything. There is only the model that fits a particular job best, and the way to find it is not guessing but testing.

What is an LLM, and why do you need to choose?

An LLM (large language model) is an AI model trained on huge amounts of text so it can understand and generate language. It is the engine behind tools like ChatGPT, Claude and Gemini. Most providers offer several sizes in the same family; Anthropic, for example, has Haiku, Sonnet and Opus, from smaller to larger, which differ in capability, speed and price. If you are new to the basics, start with AI basics: generative AI and LLMs.

Three dimensions to balance

Fast (latency)

When someone is waiting for the answer, such as a website chatbot, speed matters a lot. Smaller models usually respond faster.

Cheap (cost)

For high-volume work, such as sorting thousands of emails, small per-call prices add up to large sums. Read about cost structure in Controlling AI costs.

Smart (capability)

Complex analysis, multi-step planning or coding usually need a larger model. Anthropic's guidance notes that a more capable model often finishes a task with less work, so it can be better value despite a higher per-token price, and recommends comparing cost per completed task.

Other criteria not to forget

• Thai language: test with your real Thai text, including formal language, everyday speech and business jargon. Thailand also has models built specifically for Thai, such as Typhoon from SCB 10X.
• Context length: if you need to read long documents, check that the model can take in enough text.
• Data and server location: where is data stored, is it used for training, and does it fit the PDPA?
• Extra capabilities: such as tool calling, reading images or returning JSON.

A rough guide to matching tasks and model sizes

• Small models: classifying text, extracting data from forms, spell-checking, high-volume work whose output can be checked automatically
• Mid-sized models: drafting emails, summarising documents, answering general customer questions
• Large models: analysing multiple data sources, multi-step planning, coding, or work where mistakes are costly

This is only a starting point, not a fixed rule. The same task at two companies may need different models.

Test with your own work

Benchmarks give a broad picture, but they do not tell you how well a model will do your work. The approach Anthropic recommends is to pull sample tasks from real usage, write checks for the outputs, record cost per task next to the score, and compare several models. Before switching, run the new model in parallel on a small slice of traffic first.

Example: an online store's customer service team collects 50 real questions, tests three models, has the team lead score accuracy and tone, and looks at response time and cost too. Often a mid-sized model is enough for general questions, while complex complaints go to a larger model or a human.

Do not ask which model is the smartest. Ask which model does this job well enough at a price you can accept.

Watch out for

• New models arrive often. Test results from six months ago may no longer hold, so retest regularly.
• Prompts written for one model may not suit another. Review your prompts after switching models.
• Do not lock your system to one provider so tightly that you cannot switch. Design for swapping models from the start.

How to get started

1. Write down whether this task needs speed, price or capability first.
2. Collect 30 to 50 real task samples with good reference answers.
3. Shortlist two or three models that pass your data and language requirements.
4. Test, score and record cost per task.
5. Go live on a small scale, then review every quarter.

To learn more about AI tools that handle Thai well, read What is Claude AI, and practise writing prompts that work across models at Plants.work Learn.