AI tokens and usage limits: do we have to use it all before the reset? | Akabane Red Dev Log 16

The Snackgirls are new to both AI and coding. This dev log follows us as we try things and work through what trips us up. This time, we looked at how much AI we can use and how to avoid wasting that allowance.

Do we have to use it all just because it’s there?

Hey, it’s Red. When you work with AI, you don’t just check the results. You check how much usage you have left and when you can use it again. It’s a little funny to be checking the limit before you’ve even decided what to fix.

“Wake up… time to use those tokens.” I can see why people make that joke. But asking for random things just because we have usage left means more answers to read and sort through. So I started with what tokens are and what a reset actually restores.

Your request is input; the AI’s answer is output. Earlier messages, files it reads, and results returned by tools can also affect how much it processes. A short answer doesn’t necessarily mean the whole task was small. OpenAI usage guide · Claude usage and length limits

Nea

We should check how much material we attach, even when the question is the same.

Red

So the same number of questions could use different amounts.

“It reset!” Okay, what can we use again?

When a limit message appears, check which feature it applies to and when its allowance comes back. Not everything resets at midnight. There may be both shorter usage windows and weekly limits. Go by the time shown on your screen.

Before rushing over because someone says “They’re giving us more tokens!”, find three things in the notice: who qualifies, what increases or resets, and how long you have to use it. A higher limit, a free reset, and extra credits may all have different conditions.

Kirari

Shiro-chan, doesn’t it feel like a waste to leave some unused?

Nea

A little… but if we ask for more now, we’ll have to read those answers too.

Fair point. I decided to write down what to fix next. Being able to pick up where I left off matters more to me than filling up the usage meter.

Does switching models give us a fresh allowance?

Red

So I can switch to a lighter model? And get a fresh limit?

AIKO

Choosing a model for the task makes sense. But if the models share a limit, switching won’t erase the usage already counted.

Switching models doesn’t reset your allowance. Claude says its web app, desktop app, and Claude Code share usage limits. Codex and ChatGPT Work also have shared usage guidance. Check your service and plan to see what is counted together. Opening a new conversation doesn’t erase account usage either. Claude’s guide · Codex and Work guide

Then why switch? A tiny edit doesn’t always need the most demanding model. On the other hand, repeatedly asking a lighter model to fix a complicated bug could cost more time and usage than you meant to save.

This table is a starting point for dividing up the work. OpenAI describes Astra as a model for complex reasoning and coding, and Luna as an efficient option for tasks with a clear scope. Reconsider the model when the work changes; frequent switching isn’t a goal in itself. A lower unit price doesn’t necessarily mean fewer tokens, either. Astra · Luna’s official description

Give enough detail to avoid asking again

“Fix this.” It’s short, sure. But it doesn’t say what to fix or how. Adding the conditions one at a time later means doing the same work over and over. I’d rather include the necessary conditions and cut unrelated material.

Relevant file: [actual file path]
How to run it: [command or steps]

Change only the button labels on the start screen.
“Start” → “Start game”, “Quit” → “Exit”
Keep the button positions, colors, and click behavior unchanged.

Inspect the relevant files first and edit only what is necessary.
If possible, run the game and check the labels and button behavior.
At the end, briefly list the files changed and the checks performed.
State separately anything you couldn't verify yourself.

1. Start with relevant material. Choose what helps locate the problem, such as the error message, the affected file, and the current screen.
2. Set the output scope. Decide how much you want to read: three title options, for example, or an explanation focused on the changes.
3. Check the result before asking again. If you keep regenerating without looking, it’s hard to tell what improved.
4. Leave a handoff for the next conversation. Record the goal, current state, requirements to preserve, remaining work, and actual file locations.

A new conversation can help when moving to a new topic or cutting unnecessary long context. But if you leave out important conditions, you’ll have to explain them again. Even a summary needs checking against the actual files.

Kirari

Then I should stop asking for a hundred titles at once. I forget half of them while I’m choosing.

Red

Let’s get three, then ask for more like the one we like.

I paid for a subscription. Can there still be another charge?

Keep your included allowance, API charges, and additional usage credits separate. Paying for a subscription doesn’t mean API usage is included. If you enable extra usage, separate charges may apply after you exceed the included amount.

When I write down my next task, I’m going to add three things: what to change, which model to try first, and how to check that the fix worked. Then seeing some allowance left won’t immediately send me wondering what else I should ask for.

While Kirari packed her bag, I underlined what to check on the next edit. That’s enough for today. I’ll be back with more about making our game once I’ve gathered the actual work notes!

#VibeCoding #Devlog #AICoding #AIAutomation #BlogAutomation #AICodingforBeginners #Snackgirls #AITokens #AIUsage