AI pricing has become its own genre of confusion. Tokens, credits, messages, requests, compute units — different products use different terms for roughly the same underlying concept, making direct comparison nearly impossible. Here is a plain-English explanation of how AI pricing actually works.
What Is a Token?
A token is the basic unit of AI computation. Roughly speaking, one token equals about four characters of English text, or about 0.75 words. The sentence "Tell me about the weather" is about six tokens. A long essay might be thousands of tokens.
Language models process text in tokens — both the text you send (the prompt) and the text they generate (the completion). The total cost of any AI interaction is roughly: (tokens in your message) + (tokens in the AI's response) × (cost per token for that model).
Higher-capability models cost more per token. GPT-4o costs more than GPT-4o Mini. Claude Opus costs more than Claude Haiku. The model you use has a bigger impact on cost than the length of your messages.
What Are Credits?
Credits are a simplified abstraction over tokens. Rather than billing you for exact token counts, a service gives you a credit balance and charges some number of credits per message. This is easier to understand but hides the actual cost.
Hey Friday uses chat credits. Each message you send costs one credit, regardless of length. This simplicity is intentional: you know exactly how many messages you have left without doing math. Free accounts get 100 credits per day; premium accounts get larger monthly allocations.
Why Do Different Models Cost Different Amounts?
The cost of running an AI model scales roughly with its size. Larger models have more parameters, require more computation per token, and need more powerful (and expensive) hardware. GPT-4o, which has hundreds of billions of parameters, costs significantly more to run than GPT-4o Mini, which is designed to be efficient.
This is why many AI assistants offer multiple models at different price points. The free tier gets the smaller, cheaper model. Premium tiers unlock the more capable, more expensive models. The right choice depends on what you need: for most everyday queries, the cheaper models are completely adequate.
Subscriptions vs. Pay-As-You-Go
AI pricing generally comes in two flavors. Subscriptions give you a monthly credit or message allowance for a flat fee — predictable costs, but you may over or underuse your allocation. Pay-as-you-go charges you exactly for what you use, with no waste, but costs can be unpredictable for variable usage.
For individual consumers, subscriptions are usually better: the flat fee is worth the predictability, and most heavy users get more value than they pay for. For developers building applications, pay-as-you-go API pricing usually makes more sense.
Image Credits Are Different from Chat Credits
AI image generation is priced separately from text AI, because the computational requirements are completely different. Generating an image requires running a diffusion model, which is far more computationally intensive than generating text.
Hey Friday separates chat credits (for AI conversations) from image credits (for AI image generation). Basic image generation using Pollinations FLUX is always free. Pro-quality images use image credits, which can be earned through rewarded ads or purchased as one-time packs. Image credits never expire.
How to Maximize Your Free Allocation
- Be concise. Longer prompts use more tokens. Clear, specific questions get better answers and cost less than vague, rambling ones.
- Choose the right model. If you are asking a simple factual question, the free-tier model is usually fine. Save premium model credits for genuinely complex tasks.
- Avoid unnecessary back-and-forth. Each message in a conversation uses tokens. Include all relevant context in your initial message to reduce the number of follow-ups needed.
- Reset timing matters. If you are near your daily limit in the evening, save non-urgent queries for after the daily reset.
AI pricing will continue to fall as models become more efficient. What costs $20 per month today may cost $5 in two years. But understanding the current pricing landscape helps you make smart decisions now about which services to use and which to avoid.
Input and Output Tokens Are Not Priced the Same
This is the detail that makes AI pricing pages confusing, and it explains most surprise bills.
Providers charge separately for tokens going in and tokens coming out, and output is typically several times more expensive than input. The reason is mechanical: input can be processed in parallel, while output must be generated one token at a time, each one depending on the last. Generation is simply more expensive to compute.
The practical consequence is counter-intuitive. Pasting a long document costs less than you would expect. Asking for a long essay in reply costs more. If you are managing a budget, the lever that matters is how much you ask the model to write, not how much you give it to read.
Why a Long Conversation Gets Expensive
Here is the mechanic almost nobody explains, and it accounts for the common experience of a chat becoming slower and pricier the longer it runs.
Language models are stateless. They do not remember your previous messages. To create the impression of memory, the entire conversation so far is re-sent with every new message.
So your tenth message does not cost the price of ten words. It costs the price of the whole conversation up to that point, plus your new words. Message twenty re-sends everything from messages one to nineteen. Cost grows with the square of conversation length, not linearly.
Two things follow. Long threads get progressively more expensive and slower for the same apparent work. And starting a fresh conversation when you change topic is not merely tidy — it is materially cheaper and faster.
A Worked Example
Suppose you ask a 30-word question and receive a 300-word answer. As a rough rule, one token is about 0.75 of a word in English, so that is roughly 40 input tokens and 400 output tokens.
Now continue the conversation for ten exchanges of similar length. Your tenth question is short, but the request carries the accumulated history: roughly 4,000 tokens of prior conversation are re-sent alongside it. That single message costs around a hundred times the input of your first, for a question that looked identical in length.
This is why "clear the conversation" is the most effective cost-saving action available to you, and why services that summarise older messages rather than re-sending them verbatim can offer more generous limits.
Why Credits Exist At All
Given that providers pay per token, why do consumer services sell credits or messages instead of simply reselling tokens?
Because tokens are unpredictable in a way that makes people anxious. Nobody can estimate their monthly token consumption in advance, and a pricing model that cannot be estimated does not get adopted. A credit is a stable, comprehensible unit — one message, one image — that the provider maps onto variable underlying costs and absorbs the difference.
That abstraction is genuinely useful for users. It also means the exchange rate between credits and actual compute is set by the provider and can be adjusted. When a service quietly changes how many credits a particular model costs, that is what has happened.
Practical Rules
- Start a new chat when the topic changes. The single biggest saving available.
- Ask for the length you want. "In two sentences" genuinely costs less than letting the model decide.
- Use the cheapest model that can do the job. Most questions do not need the most capable model, and the price difference between tiers is often several-fold.
- Paste documents once, not repeatedly. Re-pasting the same text into a running conversation multiplies the cost of something already in context.