What are tokens, and why am I paying for them?
A token is the unit AI models use to measure text. It's not a word, and it's not a character—it's somewhere in between. In English, one token is roughly three-quarters of a word. So "product description" is two tokens, "hand-poured ceramic mug" is four, and a 500-word blog post runs about 650 tokens.
When you see a pricing page that says "£0.03 per 1,000 tokens," that's how the model charges you for reading your prompt and writing the answer. If you paste a 2,000-word customer email thread into Claude Sonnet 4.5 and ask it to draft a reply, you're using around 2,600 input tokens (reading the thread) plus however many output tokens the model writes (maybe 200 for a short reply). The bill is tiny—fractions of a penny—but it adds up if you're running hundreds of prompts a day.
The reason tokens matter is that they determine both cost and what you can fit into a single conversation. If you're comparing two models and one costs twice as much per token but writes tighter answers in half the length, the actual cost difference might be negligible. I learned this when I switched from a cheaper model to Gemini Flash for product descriptions. Flash cost more per token but used 30% fewer tokens per description because it didn't waffle. My monthly bill went down.
Different models count tokens slightly differently—Llama and GPT-OSS use different tokenizers than Claude—but the principle is the same. You don't need to count them yourself; every decent AI tool shows token usage after each prompt. Just check the number occasionally so you know what you're spending.
What's a context window, and why does it say my prompt is too long?
The context window is the total amount of text—input and output combined—a model can handle in one conversation. Think of it as the model's working memory. Claude Sonnet 4.5 has a 200,000-token window. Gemini Pro goes up to two million. Llama 3.3 70B sits at 128,000.
When you hit the limit, the model can't read your entire prompt. You'll get an error, or it'll silently ignore the earliest part of the conversation and give you an answer that makes no sense because it's missing context.
This comes up more often than you'd think. Say you're a sole trader running a Shopify store selling outdoor gear. You paste your entire product catalogue—300 items, each with a title, description and specs—into the prompt and ask the model to write a Boxing Day sale email. If that catalogue is 80,000 tokens and you're using a model with a 128,000-token window, you've only got 48,000 tokens left for the model's reply and any follow-up questions. The model might cut corners, or it might refuse to answer at all.
The fix is either to use a model with a bigger window (Gemini Pro is good for this) or to break the task into chunks. I once tried to feed a 15,000-word brand guideline into Claude Haiku 4.5 and got an error. I split it into three sections—tone, visuals, messaging—and ran three separate prompts. Took an extra five minutes, but it worked.
Context windows also affect multi-turn conversations. If you're refining a product description over ten back-and-forth messages, each exchange eats into the window. Eventually the model forgets what you asked three turns ago. Starting a fresh conversation resets the count.
What's an AI hallucination, and how do I stop it making things up?
A hallucination is when the model invents information that sounds plausible but isn't true. It might cite a study that doesn't exist, claim your product has a feature it doesn't, or confidently state that Royal Mail delivers on Sundays (it doesn't, except for parcels during Christmas).
This happens because AI models are pattern-matching machines, not databases. They predict what word comes next based on training data, and sometimes the most statistically likely next word is wrong. The model doesn't know it's wrong—it doesn't "know" anything. It just sounds confident.
I've seen this wreck product descriptions. A Shopify seller I know used an AI tool to write copy for a wool jumper. The model added "machine washable" because that's common in clothing descriptions. The jumper was hand-wash only. A customer bought it, machine-washed it, and left a one-star review when it shrank. The seller hadn't checked the output.
You can't eliminate hallucinations entirely, but you can reduce them. First, be specific in your prompt. Instead of "write a blog post about VAT for small businesses," say "explain the £90,000 VAT registration threshold in the UK, using only information I provide below," then paste the HMRC guidance. The model is less likely to invent details if you've given it the facts.
Second, use a model that's good at following instructions. Claude Sonnet 4.5 and DeepSeek R1 are both strong here; they're less likely to ad-lib than faster, cheaper models. Third, always fact-check anything that matters—prices, dates, regulations, product specs. If the model says your candle burns for 40 hours and you know it's 30, fix it.
What's grounding, and why do some models mention it?
Grounding is when you anchor the model's answer to a specific source of truth—a document you upload, a webpage you link, or a dataset you provide. Instead of letting the model guess, you're saying "only use this information."
This is the opposite of a hallucination. If you paste your Companies House filing into the prompt and ask the model to summarise your business structure, it's grounded in that document. It can't invent a director who doesn't exist, because the names are right there in the text.
Some tools build grounding into the interface. You upload a PDF and the model automatically treats it as the reference. In Kryotta, you can paste text directly into the prompt or use the context window to include long documents, and the model will prioritise that material over its general training data.
I use grounding every time I write product descriptions for a client. They send me a spec sheet—dimensions, materials, care instructions—and I paste it into the prompt along with "write a 100-word description using only the details in this spec sheet." The model doesn't add features that aren't there, and I don't have to spend twenty minutes fact-checking.
Grounding is especially useful for customer service. If you're drafting replies to refund requests, you can paste your returns policy into the prompt and tell the model to base its answer on that policy alone. The reply will be accurate, and you won't accidentally promise something you can't deliver.
What's the difference between open-weight and closed models, and does it matter?
A closed model is one where the company (OpenAI, Google, Anthropic) doesn't release the underlying code or weights. You can use the model through an API or a tool like Kryotta, but you can't download it, modify it, or run it on your own server. Claude Sonnet 4.5 and Gemini Pro are closed.
An open-weight model (sometimes called open-source, though that's not quite accurate) is one where the weights are public. You can download Llama 3.3 70B or Mistral Large, run them on your own hardware if you've got the specs, or use a hosted version through a platform. The company that made the model has released it for anyone to use.
For most small business owners, this distinction doesn't change your day-to-day work. You're not going to run Llama on a server in your garage. But it does affect three things: cost, privacy and flexibility.
Open-weight models are often cheaper to use because there's more competition among hosting providers. Closed models can be more expensive because one company controls access. If you're running hundreds of prompts a day, that adds up.
Privacy is the second factor. With a closed model, your prompts go through the company's servers. Most providers say they don't train on your data, but you're trusting their word. With an open-weight model hosted by a privacy-focused provider—or run on your own hardware if you're technical—you've got more control. This matters if you're handling customer data or drafting sensitive business documents.
Flexibility is the third. Open-weight models can be fine-tuned (retrained on your specific data) if you need something highly specialised. A closed model is what it is. For most use cases, though, the closed models are more capable out of the box, so unless you've got a very niche need, you're fine sticking with Claude or Gemini.
Why does any of this matter when I just want to write an email?
Because understanding these terms stops you wasting money and getting frustrated. If you don't know what a context window is, you'll keep pasting 10,000-word documents into a model that can't handle them, then blame the AI when it gives you nonsense. If you don't know what a hallucination is, you'll publish product copy that claims your soap is vegan when it contains beeswax, and you'll deal with the fallout.
These aren't academic concepts. They're the reason one model costs £15 a month and another costs £150. They're why your prompt works beautifully in Claude Sonnet but fails in a cheaper alternative. They're why your competitor's AI-generated product photos look real and yours look like clipart.
You don't need to become a machine-learning engineer. But knowing that tokens measure usage, context windows set limits, hallucinations are predictable, grounding prevents invention, and open-weight models offer different trade-offs—that's enough to make better decisions. You'll pick the right model for the job, write prompts that actually work, and catch mistakes before they reach customers.
I've watched sole traders spend hours fighting with AI tools because they didn't understand why the model kept "forgetting" the start of their conversation (context window) or why it invented a feature their product didn't have (hallucination). Ten minutes learning this jargon would've saved them the headache.
Questions people ask
Do I need to count tokens manually before running a prompt?
No. Every decent AI tool shows you token usage after you submit the prompt—input tokens, output tokens and total cost. Just glance at the number occasionally so you know what you're spending. If a task is using way more tokens than expected, you can simplify the prompt or switch to a cheaper model for routine work.
Can I stop a model from hallucinating completely?
Not entirely, but you can make it much less likely. Be specific, provide source material in the prompt (grounding), and use a model known for accuracy like Claude Sonnet 4.5 or DeepSeek R1. Always fact-check anything that matters—prices, regulations, product claims.
If I hit the context window limit, will the model tell me or just give a bad answer?
It depends. Some models throw an error and refuse to run. Others silently truncate the input and answer based on whatever fits, which is worse because you won't realise it's missing half your prompt. If your task involves long documents, check the model's context window before you start.
Should I use an open-weight model if I'm worried about privacy?
If you're handling sensitive customer data or confidential business documents, an open-weight model hosted by a privacy-focused provider gives you more control. For general marketing copy or product descriptions, a closed model like Claude or Gemini is fine—most providers don't train on your data, and the quality is often higher.
If you're comparing models or just trying to understand what you're paying for, Kryotta gives you access to Claude, Gemini, Llama, Mistral, DeepSeek and more in one workspace, with token counts and costs shown clearly after every prompt. You can test a prompt across multiple models in Compare, or stick with one once you know what works. Start a free trial and see which model fits your workflow—no jargon required.
