If you pay for AI by credits or points, you have probably seen it: the first few messages of a chat cost almost nothing, then the same kind of question starts costing three, five, ten times as much. People who move between services, Poe users in particular, often describe a monthly allowance that should have lasted the month running out in under two weeks.
Nothing is broken. It is how language models are billed, and once you see the arithmetic you can control it. This guide explains why it happens, shows real numbers for two models on Essevin, and lists the habits that make credits last.
The short answer
A language model does not remember your chat. Every time you send a message, the app sends the whole conversation so far back to the model, plus your new message. You pay for everything the model reads (input) and everything it writes (output). So the cost of a message grows with the length of the conversation behind it.
Why every message re-sends the whole chat
Models from Anthropic and OpenAI are billed in tokens, roughly three quarters of an English word or one to two Chinese characters. A reply only makes sense if the model sees what came before, so each request carries the full history. A chat with twenty back-and-forth turns means the twentieth request carries all nineteen earlier questions and answers.
Services that sell credits on top of these models, Essevin and Poe included, usually pass that structure through. The points or credits a message uses then depend on the model and on how much text the request carries. On Essevin, credits follow the official model price times a fixed credit rate, and 10,000 credits = HK$1, so you can turn any request into Hong Kong dollars.
What it looks like in numbers
The table assumes each answer is about 800 tokens long and that every turn adds roughly 1,000 tokens to the history. Figures are Essevin credits per message, rounded up, without any caching discount.
| Conversation behind the message | Claude Opus 5.5 | GPT-6.1 Sol |
|---|---|---|
| Fresh chat (about 3,000 tokens) | about 1,771 credits (HK$0.18) | about 322 credits (HK$0.03) |
| About 10 turns in (10,000 tokens) | about 3,541 credits (HK$0.35) | about 644 credits (HK$0.06) |
| About 20 turns in (20,000 tokens) | about 6,070 credits (HK$0.61) | about 1,104 credits (HK$0.11) |
| About 40 turns in (40,000 tokens) | about 11,129 credits (HK$1.11) | about 2,023 credits (HK$0.20) |
Two things stand out. The fortieth message on the same model costs more than six times the first. And the choice of model matters even more than length: a long GPT-6.1 Sol chat is still cheaper per message than a fresh Claude Opus 5.5 one.
Five habits that make credits last
Start a new chat when the topic changes. This is the biggest saving. A new chat drops the history you no longer need.
Carry a summary, not the whole thread. Before you leave a long chat, ask for a five-line summary of what was decided, then paste it into a fresh chat.
Match the model to the job. Use GPT-6.1 Sol or Claude Haiku 5.5 for drafting, translating and quick questions. Save Claude Opus 5.5 for the hard part: long documents, careful reasoning, difficult code.
Attach only what the model needs. A whole 60-page PDF is read again on every turn. Paste the section you are asking about.
Check what each request used. Look at the usage record after a session. Patterns show up fast.
How Essevin handles this
Essevin is a monthly subscription priced in Hong Kong dollars. Each month you get one pool of credits for chat, images and video, and credits reset every month and do not carry over. Long chats cost more per message here too, so the habits above apply. What Essevin adds is a ceiling and a record:
A per-message cap. Every plan sets the most one message can cost: 10,000 credits (HK$1) on Taster, 30,000 on Everyday, 60,000 on Advanced, 120,000 on Professional. Before a message is sent, Essevin works out the most it could cost. The reply length is limited so it stays inside the cap, and if the conversation alone is too long to fit, the message is refused with a notice instead of being charged. The trade-off is real: on an expensive model, near the cap, answers can be shorter. That is the moment to start a new chat or switch model.
A usage record. The console shows what every request used, so you can see which chats are expensive.
Published estimates. The plans page and models page give credits per standard message for each model, so you can size a plan before you pay.
FAQ
Why did the same question cost more today than yesterday?
Most likely it was asked deeper into a longer conversation, or on a more expensive model. The question itself is a small part of what the model reads; the history behind it is usually the larger part.
Does starting a new chat lose anything?
Only the context. The model will not know what you discussed before unless you tell it. A short summary pasted at the start of the new chat keeps the useful part.
How many messages does a monthly plan give me?
It depends on model and length. Essevin publishes estimates for a standard message of 3,000 input and 800 output tokens: the HK$28 Taster plan is about 776 GPT-6.1 Sol messages or about 141 Claude Opus 5.5 messages a month, and the HK$128 Everyday plan about 3,105 or about 564. See the plans page for every plan.
Is a per-message cap good or bad?
Both. It means no single message can surprise you with a large charge. It also means very long chats on expensive models can hit the cap and get shorter answers. Starting a new chat fixes it.
Do credits carry over if I don't use them?
No. On Essevin, monthly credits reset every month and do not carry over. Optional credit packs last the month you buy them and the next. Details are in the subscription guide.
Want to see your own numbers? Create an account, compare the plans, and check the usage record after your first session.