How Far a Free AI Allowance Actually Goes
- Free AI access is metered two ways: a count of messages, or an allowance of tokens. Tokens are what most systems really track.
- A token is about three quarters of a word in English, and your question and the answer both get counted.
- On askai.free you get one question with no signup at all, then a free account with 100,000 starter tokens.
- One quick question with a short answer costs roughly 230 tokens, so that allowance covers about 430 of them.
- Long threads are the expensive part. The model re-reads every earlier message on every turn, so turn 20 can cost twenty times what turn 1 cost.
You want the number before you spend any attention. Most sites answer with a soft word like "generous" and hope you stop asking. The real answer is arithmetic, and it takes about a minute to learn.
The number on askai.free, stated plainly
One question needs no signup at all. No email, no card, nothing. After that you make a free account, and it comes with 100,000 starter tokens.
Free accounts can use ChatGPT 4o, Claude Sonnet 3.5 and Gemini Flash. Those tokens are spent across all of them, from one pot. When the pot is empty, Pro at $14.99 a month refills it with 25,000,000 tokens.
That's the whole offer. The rest of this page explains what those numbers buy. A figure like 100,000 means nothing until you can turn it into work you actually do.
The two counting systems, and why sites pick one
Message counts
Some products give you a flat number of messages. Ten a day, thirty a month, five before the wall. It's easy to understand and easy to display.
It's also crude. A message asking "capital of Peru" and a message containing your entire dissertation cost the same under this system. Providers who use it usually have to add a second, hidden limit to stop the dissertation case. That's why you sometimes hit a wall well before your counter says you should.
Token allowances
Most systems really meter tokens, because that's what the model actually charges for underneath. An allowance flexes with the size of the job. Short questions barely dent it, and huge documents eat it fast.
Plenty of sites run both at once. On askai.free, Pro carries 500 questions a month and 25M tokens. Whichever runs out first stops you. Worth checking on any plan you compare, because the smaller of the two numbers is your real limit.
What a token is, without the jargon
Models don't read words. They read tokens, which are chunks of text roughly three to four characters long. Common short words are one token each. Longer words split up, so "tokenization" becomes about four pieces.
The rule of thumb worth memorising: one token is about 0.75 of a word. Flip it around and 100 words is about 133 tokens. Multiply your word count by 1.33 and you're close enough for planning.
Here's the part people miss. Both sides get counted. Your question is input tokens, the reply is output tokens, and your allowance is charged for the pair. Since the reply is usually far longer than the question, the model's own writing is where most of your allowance goes. The Token glossary entry has a longer conversion table if you want to check the maths.
Two worked examples you can copy
A short question with a medium answer
- Your question, 20 words, about 30 tokens.
- The answer, 300 words, about 400 tokens.
- Total for the exchange: around 430 tokens.
Do that 230 times and you've used roughly 100,000. Ask shorter questions and cap the answers and you'll get more.
A long document with a long answer
- A 2,000 word report pasted in, about 2,670 tokens.
- Your instruction on top, 30 words, about 40 tokens.
- The answer, 800 words, about 1,070 tokens.
- Total for the exchange: around 3,800 tokens.
One document job costs about nine quick questions. Nothing has gone wrong there. That's simply what a bigger job weighs.
Why turn 20 costs twenty times what turn 1 did
This is the single most useful thing most people never get told. A model has no memory between turns. To keep the thread coherent, the product re-sends the whole conversation every single time you hit send.
So your fifth question isn't really your fifth question. It's the whole transcript so far, plus your fifth question, handed over again from the top.
Watch it compound. Say every question is 40 tokens and every answer is 400, so each exchange adds 440 tokens to the transcript.
- Turn 1 costs about 440.
- Turn 2 costs about 880, because turn 1 gets re-read.
- Turn 3 costs about 1,320.
- Turn 20 costs about 8,800 on its own.
Add every turn together and a 20 turn thread runs to roughly 92,400 tokens. That's almost the entire free allowance in one conversation.
Now the comparison that matters. Twenty separate questions, each in a fresh chat, cost about 8,800 in total. Same twenty questions, same answers, one long thread: about 92,400. Ten times the cost for the same work.
Long threads are still worth it when the model genuinely needs what came before. Editing a draft, debugging one problem, planning one trip. They're pure waste when you've moved on to something unrelated and never opened a new chat. Some products trim old messages once a thread gets huge. That caps the damage. It also means the model quietly forgets your early instructions. The Context window glossary entry covers where that ceiling sits.
What each job costs, and how far 100,000 goes
| What you're doing | Rough tokens per go | Times before 100,000 runs out |
|---|---|---|
| Quick question, short answer | ~230 | ~430 |
| Rewrite a short email | ~550 | ~180 |
| Ask for a full explanation of a topic | ~740 | ~135 |
| Summarise a 2,000 word document | ~3,100 | ~32 |
| Turn 1 of a chat thread | ~440 | ~227 |
| Turn 20 of that same thread | ~8,800 | ~11 |
| A whole 20 turn thread, start to finish | ~92,400 | ~1 |
Read the bottom three rows together. Identical questions, wildly different prices, and the only variable is how long the thread had grown.
Example figures, to show the shape of the trade-off. Built from the 0.75 words per token rule and the typical answer lengths described above, then divided into the 100,000 token free allowance. Your real usage will differ with answer length and model. Bars are scaled to the largest value, 430.
Five habits that stretch a free allowance
None of these change the answer you get. They change how much of your allowance the answer costs.
Two mistakes that burn tokens for nothing
The first is the endless thread. People keep one chat open for weeks, asking about dinner, then a bug, then a birthday card. Every unrelated turn drags the whole history along with it.
Mistake two is asking a model to repeat work it already did. "Write it again but shorter" re-reads the long version, then writes a new one. That's cheaper than starting from scratch, and dearer than asking for 150 words the first time.
One more thing to know about free tiers generally. On askai.free, free and anonymous chats are deleted after 24 hours. Copy anything you want to keep before you close the tab. The AskAI.free FAQ covers what's stored, for how long, and what happens when you upgrade.
Work out your own number before you sign up anywhere
Take the job you actually do most. Count the words you'd paste in, and the words you'd want back. Add them, multiply by 1.33, and you have your cost per go. Divide any free allowance by that and you'll know exactly how far it stretches, on any site, not just this one.
Then test it. Ask your real question in the free AI Chat and see what a genuine answer costs you. Putting the same prompt to two models is worth doing too. A chattier model spends more of your allowance on the same question, and the side by side pages in Compare AI Models show that gap in one screen.
Free is enough for a real trial. It's not enough for daily document work, and now you can prove which one you are with arithmetic rather than a guess.