Tokens: the meter running in the background
A token is a piece of a word. "Running" is one token. "Run" is one token. "Unbelievable" is usually two. Every time you send a message to an AI model and it writes back, you're using tokens—your input costs tokens, the response costs tokens, and if you've pasted a long product description or a customer email thread for context, all of that costs tokens too.
Why it matters: most AI platforms charge by the token, and if you're on a subscription with a monthly cap, you hit that cap faster when you paste entire email threads or ask the model to rewrite the same caption five times. On Kryotta, Claude Sonnet 4.5 is more expensive per token than Gemini Flash, so if you're generating fifty Instagram captions in one sitting, switching models can cut your bill in half.
Here's what that looks like in practice. I run a small jewellery store on Etsy. I was using Claude to write all my product descriptions—about 120 words each, fifteen new listings a week. One month I noticed I'd burned through my token allowance by the 18th. I switched to Gemini Flash for the first draft of each description, then used Claude only for the final polish on the listings I cared about most (the $80+ pieces that needed a warmer, more personal voice). My bill dropped by 40 per cent and the descriptions still converted.
The short version: if you're doing high-volume work—bulk product uploads, fifty email replies a day, twenty TikTok scripts in one session—pick a cheaper, faster model for the first pass. Save the expensive model for the work that actually makes you money.
Context windows: how much the model remembers
The context window is how much text the model can hold in its head at once—your prompt, the conversation so far, any documents you've pasted in, and the space left over for the response. Gemini Flash has a one-million-token context window. Claude Sonnet 4.5 has 200,000. GPT-4o has 128,000. When you hit the limit, the model forgets the beginning of the conversation and starts giving answers that contradict what it said three messages ago.
Why it matters: if you're pasting a full Shopify order export to ask "which customers bought twice in November but haven't come back," or if you're feeding the model a fifty-message customer service thread to draft a refund email, you need a big context window. If you pick a model with a small window, it'll truncate the data and give you an answer based on incomplete information.
I learned this the hard way during Black Friday last year. I manage Instagram and TikTok for a small activewear brand—three team members, about $40,000 a month in revenue. We had 900 customer emails in the inbox by Saturday morning, and I wanted to sort them by urgency: refunds first, shipping questions second, general inquiries last. I pasted the whole export into ChatGPT and asked it to tag each one. It worked for the first 200 emails, then started repeating the same three tags and missing obvious refund requests. The context window had filled up, the oldest emails had dropped out of memory, and the model was now guessing.
I switched to Gemini Pro, which has a much larger window, re-pasted the export, and got accurate tags for all 900 messages. Took fifteen minutes instead of two hours of manual sorting.
The short version: if you're working with long documents, big data exports, or multi-hour chat threads, check the context window before you start. A model with a small window will forget half your input and give you useless answers.
Hallucinations: when the model just makes things up
A hallucination is when the model states something as fact that isn't true. It might invent a Stripe policy that doesn't exist, cite a customer review you never received, or generate a product specification with the wrong measurements. It's not lying—it's pattern-matching from its training data and filling in gaps with plausible-sounding text that happens to be wrong.
Why it matters: if you're using AI to write product descriptions, answer customer questions, or draft refund policies, a hallucination can cost you money. A wrong dimension on an Etsy listing means returns. A fake Stripe fee in a customer email means confusion and lost trust. You can't just copy-paste AI output and assume it's accurate.
Here's what I do now. I sell vintage furniture on Shopify and eBay. I used to ask Claude to write descriptions based on photos—"mid-century teak sideboard, three drawers, brass handles, 60 inches wide." One day a customer messaged me: "You said 60 inches, but my tape measure says 54." I'd never measured it myself. I'd trusted the AI's guess from the photo, and the AI had hallucinated a plausible width based on similar sideboards in its training data.
Now I give the model exact measurements, materials and condition notes in the prompt, and I tell it: "Do not add any details I haven't provided. If I haven't mentioned the wood type, say 'wood' not 'teak.'" I also spot-check every fifth description against my notes. It takes an extra two minutes per listing, but I haven't had a complaint since October.
The short version: never trust an AI model to know facts it wasn't explicitly told. If it's important—measurements, prices, policies, dates—check it yourself before you publish.
Grounding: giving the model something true to work from
Grounding means anchoring the AI's response to a specific source of truth—a document, a dataset, a URL, a customer email. Instead of asking "What's my Etsy refund policy?", you paste your actual policy and ask "Rewrite this policy in simpler language for my FAQ page." The model can't hallucinate because it's working from your text, not from its training data.
Why it matters: grounding eliminates most hallucinations and makes the output more accurate. It's especially useful for customer service replies, product descriptions based on spec sheets, and any task where you already have the information but need it rewritten or summarised.
I run a small candle business on Shopify. I get the same five customer questions every week: shipping times, ingredient lists, burn hours, wholesale pricing, custom orders. I used to ask Claude "Write a reply to a customer asking about shipping times," and the model would generate something generic that didn't match my actual shipping policy. Now I keep a Google Doc with my real answers—two-day processing, USPS Priority, tracking included for orders over $50—and I paste the relevant section into the prompt every time. The model rewrites it in a friendly tone, matches the customer's level of urgency, and never invents a shipping option I don't offer.
Grounding also works for product descriptions. If you're selling on Etsy and you have a supplier spec sheet—dimensions, materials, care instructions—paste that into the prompt and ask the model to turn it into a description. You'll get accurate details in a voice that sounds like you, and you won't have to fact-check every claim.
The short version: if you already have the information, paste it into the prompt. The model will rewrite it, expand it, or summarise it, but it won't make things up.
Open-weight vs closed models: who can see how it works
A closed model is owned by one company, and the weights—the internal settings that make the model work—are secret. GPT-4o, Claude, Gemini: all closed. You use them through an API or a web interface, and you have no idea how they're built or what data they were trained on. An open-weight model publishes its weights. Anyone can download them, run the model on their own hardware, and see exactly how it works. Llama 3.3 70B, Mistral Large, DeepSeek V3: all open-weight.
Why it matters: closed models are usually more polished and better at complex tasks, but you're locked into one company's pricing and terms. Open-weight models give you more control—you can run them locally if you have the hardware, or you can use a platform like Kryotta that hosts them for you—and the pricing is often lower because there's competition. For most small business work, the difference in quality is small enough that the cost savings matter more.
I manage TikTok and Instagram for a small home-goods brand. We post three times a week, and I use AI to write captions, generate product carousels, and draft replies to DMs. I started with ChatGPT Plus at $20 a month, but I was hitting the rate limit during busy weeks and the cost per caption felt high. I switched to Kryotta and started using Llama 3.3 70B for captions and DeepSeek V3 for longer scripts. Both are open-weight, both are fast, and the quality is close enough to GPT-4o that my engagement rate didn't drop. My monthly AI bill went from $20 to $11, and I stopped worrying about rate limits.
The trade-off: closed models like Claude are sometimes better at tone and nuance, especially if you're writing customer service emails or anything that needs warmth. I still use Claude for refund replies and sensitive messages. But for high-volume work—fifty captions, twenty product descriptions, a hundred email subject lines—open-weight models are faster and cheaper, and the output is good enough.
The short version: closed models are slightly better at the hard stuff, but open-weight models are cheaper and faster for repetitive work. Use both, and pick the model that fits the task.
Questions people ask
Do I really need to understand tokens if I'm just writing Etsy descriptions?
You don't need to count every token, but knowing that longer prompts cost more helps you decide when to trim a pasted email thread or switch to a cheaper model. If you're doing fifty descriptions a week, small token savings add up.
How do I know if a model is hallucinating?
Spot-check any claim you didn't provide. If you gave the model a photo and it added a material, measurement or feature you didn't mention, that's a hallucination. Ground your prompts with real data and the problem mostly goes away.
Is an open-weight model safe to use for customer emails?
Yes. Open-weight doesn't mean unmoderated or risky—it just means the model's weights are public. Llama, Mistral and DeepSeek are all used by large companies for customer-facing work. Just check the output before you send it, same as you would with any model.
Which model should I start with?
For product descriptions and captions, try Gemini Flash or Llama 3.3 70B—they're fast and cheap. For customer service emails or anything that needs a warm tone, use Claude Sonnet 4.5. If you're not sure, Kryotta lets you compare models side by side on the same prompt, so you can see which one fits your voice before you commit.
You don't need a computer science degree to use AI well. You need to know what tokens cost, what the model remembers, when it's guessing, and how to stop it from making things up. Those five concepts—tokens, context windows, hallucinations, grounding, open-weight vs closed—cover 90 per cent of the decisions you'll make when you're writing descriptions, answering emails, or drafting a Black Friday campaign. The rest is just practice.
If you want to try the multi-model approach without juggling subscriptions, Kryotta gives you Claude, Gemini, Llama, Mistral, DeepSeek and the rest in one workspace for $14 a month. Pick the model that fits the task, compare outputs side by side, and stop worrying about rate limits during a busy week.
