--- collection: lessons slug: choosing-an-ai-model locale: en title: "Why so many AI models?" summary: "The same Claude question produced seven trips from Haiku and one from Sonnet: learn to choose a model by task, speed and cost, then check the result yourself." publishedAt: 2026-10-18 status: live tags: ["Model","Claude","Benchmarks","Tokens"] canonical: https://codebelkalam.com/en/learn/choosing-an-ai-model --- # Why so many AI models? The same Claude question produced seven trips from Haiku and one from Sonnet: learn to choose a model by task, speed and cost, then check the result yourself. [▶ watch](https://youtube.com/watch?v=Pqa1ht8ioFY) > **Note:** > > The video is in Arabic. This lesson explains the experiment and its takeaways in English. The translated prompts below let you try the idea in English; changing the language may change the results. ## The short version You open an AI app and find a list of models. Do you choose the most famous, the strongest, or the fastest? Start with the job you need done, rather than the model's name. In the episode, we give two models inside Claude exactly the same question, each in a fresh chat. **Haiku** answers with seven trips; **Sonnet** answers with one. The difference is how they handle a condition in the question. The experiment helps explain model differences, the trade-offs between capability, speed and cost, and why you still need to check the result yourself. ## Key terms - **Model**: the system behind the chat app that processes your request and generates a response. - **Chatbot**: the interface through which you talk to a model; it may let you choose among several models. - **Input Tokens**: the units the model reads in your request and the context sent with it. - **Output Tokens**: the units the model generates in its response. - **Frontier Models**: models that push the limits of available capabilities at the time of release; the label does not guarantee a correct answer every time. - **Benchmarks**: sets of standardized tests used to compare model performance on particular tasks. ## The explanation ### Same counter, different clerk Imagine a government office: the counter is the **Chatbot**, and the clerk behind it is the **Model**. You hand over the same paper through the same counter, but a different clerk may receive it. A less experienced clerk might handle clear, repetitive requests quickly and at a lower cost. A more experienced clerk might handle conditions and exceptions better, but their work may take longer and cost more. This is an analogy for differences in capability, speed and cost, rather than a literal explanation of how models work. That is why one app can offer several model names and versions. Every task does not need the strongest available model. ### The boat experiment: one condition changes the answer The recorded prompt was in Egyptian Arabic. Here is an English translation: ```prompt A farmer has a wolf, a goat and a cabbage, and wants to cross a river. The boat is large enough to carry him, the wolf, the goat and the cabbage all together at once. What should he do to get them all across in the fewest trips? ``` In the recording, Haiku gave the familiar puzzle solution: seven trips, transporting the items in stages. Sonnet noticed the changed condition: **the boat can carry everyone at once**. They all board and cross in one trip. Seven trips answer a different version of the puzzle, with a small boat. They do not answer the question written here. You do not need programming experience to check this: read the boat's capacity, then compare it with the proposed solution. > **Warning:** > > This does not establish that Sonnet is always correct or Haiku is unsuitable. It is one question in one recording; results may change when you repeat it, change the task, or use another version. Model access on free plans at recording time also does not guarantee access today. ### Why cost and speed matter Remember that text is split into **Tokens**. In API pricing, providers list an input price and an output price, and rates can differ between models. Compare both prices using the same unit, and consider the length of your request and response as well as the model name. This is different from a chat app's subscription price. Usage limits and waiting time can also affect your choice, even when you do not pay for each request individually. Do not assume a subscription includes API credit; check the terms of the service you use. We do not freeze a price list or model ranking here: versions, prices and availability change. Check [Anthropic's models](https://platform.claude.com/docs/en/models/overview) and [API pricing](https://platform.claude.com/docs/en/about-claude/pricing), or [OpenAI's models](https://developers.openai.com/api/docs/models) and [API pricing](https://developers.openai.com/api/docs/pricing). ### What do benchmarks tell you? **Benchmarks** help compare models on the tests they were given. They do not promise that the top-scoring model will do your particular job the way you want. Look at the test and its conditions, then try a task that resembles your work. The model race in the episode is illustrative, rather than an actual ranking or published test result. ## Try it yourself ### 1. Choose two models you can access, and open a fresh chat for each. Record the displayed name and version, and the date. Keep instructions and tools as similar as you can. If you cannot choose a model manually, record that instead of guessing its name. ### 2. Send the boat question above to both, using identical wording and language. Save the original replies. Do not change the question between attempts or keep retrying until you get the winner you prefer. ### 3. Check both replies yourself: did each notice that everyone fits in the boat? Did it give the fewest trips? Record the specific mistake, rather than just your impression of the writing style. ### 4. Try a task close to your own work, with clear conditions you can check, such as the request below. Send the same text in two fresh chats. ```prompt Write a formal email to my manager asking to move the report deadline from Tuesday to Thursday because I am waiting for sales data. Use three lines, and do not add a name, date or reason I have not supplied. ``` ### 5. Compare adherence to the conditions, waiting time, and how much editing you needed. Open the official pricing pages and compare Input Tokens and Output Tokens using the same unit. Write down why you would choose a model for this type of task, then try other tasks before generalizing. ## Why this matters for building software When you build an app by talking to AI, your requests vary in difficulty. Planning, writing code, changing a button label, and investigating a bug are different jobs. You can start with a fast model for a simple, checkable task, then try a more capable model when conditions or exceptions are not handled well. The episode's division of work—Opus for planning and review, Sonnet for implementation, and Haiku for repetitive edits—is an example, rather than a fixed rule or a promise that Claude Code or Codex automatically routes work that way. Choose for the task, and check the result. For software, another model's review is not enough: run the app and test that the requested behavior actually works. ## Related lessons - [What is AI?](/en/learn/what-is-ai): the difference between a Model and a Chatbot. - [How does AI think?](/en/learn/what-is-a-token): Tokens, input and output. - [How do you ask for exactly what you want?](/en/learn/what-is-a-prompt): writing a clear request with checkable conditions. - [Why does AI forget what you told it?](/en/learn/context-and-memory): what goes into a chat's Context. - [Why does AI lie so confidently?](/en/learn/what-is-hallucination): why a confident tone does not establish correctness.